In [1]:
import os
import cv2
import glob
import pylab
import keras
import random
import warnings
import shutil
import numpy as np
import splitfolders
import pandas as pd
import seaborn as sns
from PIL import Image
import pydicom as dicom
import tensorflow as tf
%matplotlib inline
warnings.simplefilter("ignore")
import matplotlib.pyplot as plt
from skimage.transform import resize
from keras.preprocessing import image
from matplotlib.patches import Rectangle
from keras.applications.vgg16 import VGG16
from keras.optimizers import Adam,SGD,RMSprop
from keras.models import Sequential,load_model,Model
from sklearn.model_selection import train_test_split
from keras.applications.vgg16 import preprocess_input
from keras.preprocessing.image import ImageDataGenerator
from sklearn.metrics import classification_report,confusion_matrix
from keras.callbacks import ModelCheckpoint, ReduceLROnPlateau
from keras.layers import GlobalAveragePooling2D, GlobalMaxPooling2D, TimeDistributed
from keras.layers import Dense, Conv2D , MaxPool2D , Flatten , Dropout , BatchNormalization,Layer,Input,InputSpec

from keras import backend as K
from keras.utils.layer_utils import get_source_inputs
from keras.utils import layer_utils
from keras.utils.data_utils import get_file
#from keras.objectives import categorical_crossentropy
from keras.utils import generic_utils
from keras import initializers, regularizers
In [3]:
#!pip install -U scikit-learn

LABEL DATA¶

In [4]:
print("==================================")
label_data=pd.read_csv('stage_2_train_labels.csv')
class_info=pd.read_csv('stage_2_detailed_class_info.csv')
print("Total records in label_data : ",len(label_data))
print("label_data Target value counts : ")
print(label_data['Target'].value_counts())
print("==================================")
print("Total records in class_info : ",len(class_info))
print("class_info class value counts : ")
print(class_info['class'].value_counts())
print("==================================")

==================================
Total records in label_data :  30227
label_data Target value counts : 
0    20672
1     9555
Name: Target, dtype: int64
==================================
Total records in class_info :  30227
class_info class value counts : 
No Lung Opacity / Not Normal    11821
Lung Opacity                     9555
Normal                           8851
Name: class, dtype: int64
==================================

We have equal number of data rows in label_data & class_info ie 30277

Above value counts show that we have:¶

non-pneumonia cases i.e "0" : 20672 in the data label

pneumonia cases i.e "1" : 9555 in the data label

In [5]:
# merging the data from class info and labels using the patientid
df_data = pd.merge(left = class_info, right = label_data, how = 'left', on = 'patientId')
In [6]:
#checking how duplicates exists in the data
print(sum(df_data.duplicated()))

7402

After merging the data we found out 7402 records are duplicate and hence they can be removed¶

In [7]:
df_data.drop_duplicates(inplace= True)
In [8]:
df_data.shape
Out[8]:
(30227, 7)
In [9]:
df_data.info()

<class 'pandas.core.frame.DataFrame'>
Int64Index: 30227 entries, 0 to 37626
Data columns (total 7 columns):
 #   Column     Non-Null Count  Dtype  
---  ------     --------------  -----  
 0   patientId  30227 non-null  object 
 1   class      30227 non-null  object 
 2   x          9555 non-null   float64
 3   y          9555 non-null   float64
 4   width      9555 non-null   float64
 5   height     9555 non-null   float64
 6   Target     30227 non-null  int64  
dtypes: float64(4), int64(1), object(2)
memory usage: 1.8+ MB
In [10]:
df_data.head()
Out[10]:
patientId class x y width height Target
0 0004cfab-14fd-4e49-80ba-63a80b6bddd6 No Lung Opacity / Not Normal NaN NaN NaN NaN 0
1 00313ee0-9eaa-42f4-b0ab-c148ed3241cd No Lung Opacity / Not Normal NaN NaN NaN NaN 0
2 00322d4d-1c29-4943-afc9-b6754be640eb No Lung Opacity / Not Normal NaN NaN NaN NaN 0
3 003d8fa0-6bf1-40ed-b54c-ac657f8495c5 Normal NaN NaN NaN NaN 0
4 00436515-870c-4b36-a041-de91049b9ab4 Lung Opacity 264.0 152.0 213.0 379.0 1

Writing a function that can tell us about missing data in the dataframe of labels¶

In [11]:
# Check for missing data
def check_for_missing_data(df):
    total = df.isnull().sum().sort_values(ascending=False)   # total number of null values
    percent = (df.isnull().sum()/df.isnull().count()).sort_values(ascending=False) #percentage of values that are null
    missing_data = pd.concat([total, percent], axis=1, keys=['Total', 'Percent']) # putting the above two together
    return missing_data
In [12]:
#for labels
check_for_missing_data(label_data)
Out[12]:
Total Percent
x 20672 0.683892
y 20672 0.683892
width 20672 0.683892
height 20672 0.683892
patientId 0 0.000000
Target 0 0.000000
In [13]:
check_for_missing_data(class_info)
Out[13]:
Total Percent
patientId 0 0.0
class 0 0.0

There are no missing value(s) found in the class_info table and all the target values are filled therefore there is no empty value, cells that have Nan or empty cell are having no phenumonia so we should replace those value with 0¶

In [14]:
label_data.fillna(0, inplace=True)
In [15]:
label_data.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 30227 entries, 0 to 30226
Data columns (total 6 columns):
 #   Column     Non-Null Count  Dtype  
---  ------     --------------  -----  
 0   patientId  30227 non-null  object 
 1   x          30227 non-null  float64
 2   y          30227 non-null  float64
 3   width      30227 non-null  float64
 4   height     30227 non-null  float64
 5   Target     30227 non-null  int64  
dtypes: float64(4), int64(1), object(1)
memory usage: 1.4+ MB
In [16]:
label_data[label_data["Target"]==1].describe().T
Out[16]:
count mean std min 25% 50% 75% max
x 9555.0 394.047724 204.574172 2.0 207.0 324.0 594.0 835.0
y 9555.0 366.839560 148.940488 2.0 249.0 365.0 478.5 881.0
width 9555.0 218.471376 59.289475 40.0 177.0 217.0 259.0 528.0
height 9555.0 329.269702 157.750755 45.0 203.0 298.0 438.0 942.0
Target 9555.0 1.000000 0.000000 1.0 1.0 1.0 1.0 1.0

from above we can get an idea of mean width and mean height of the phenumoniatic image areas¶

In [17]:
print(label_data["Target"].value_counts())
np.unique(label_data["Target"])

0    20672
1     9555
Name: Target, dtype: int64
Out[17]:
array([0, 1], dtype=int64)

Pie chart to show the % of class label data¶

In [18]:
#sns.countplot(label_data["Target"].value_counts())
df_data["class"].value_counts().plot(kind='pie',autopct='%1.1f%%')
Out[18]:
<AxesSubplot: ylabel='class'>

Pie chart to show the % of Target label data¶

In [19]:
#sns.countplot(label_data["Target"].value_counts())
df_data["Target"].value_counts().plot(kind='pie',autopct='%1.1f%%')
Out[19]:
<AxesSubplot: ylabel='Target'>
In [20]:
#No Lung Opacity / Not Normal has been classified as Target 0
df_data[df_data["patientId"]=='0004cfab-14fd-4e49-80ba-63a80b6bddd6']
Out[20]:
patientId class x y width height Target
0 0004cfab-14fd-4e49-80ba-63a80b6bddd6 No Lung Opacity / Not Normal NaN NaN NaN NaN 0
In [21]:
#Normally classified data sample with Target 0
df_data[df_data["patientId"]=='003d8fa0-6bf1-40ed-b54c-ac657f8495c5']
Out[21]:
patientId class x y width height Target
3 003d8fa0-6bf1-40ed-b54c-ac657f8495c5 Normal NaN NaN NaN NaN 0
In [22]:
#Pneumonia case classified with target as 1b
df_data[df_data["patientId"]=='00436515-870c-4b36-a041-de91049b9ab4']
Out[22]:
patientId class x y width height Target
4 00436515-870c-4b36-a041-de91049b9ab4 Lung Opacity 264.0 152.0 213.0 379.0 1
5 00436515-870c-4b36-a041-de91049b9ab4 Lung Opacity 562.0 152.0 256.0 453.0 1

Lung Opacity has been classified as 1 and other classes has been classified as Normal indicated as 0¶

The class value 'No Lung Opacity / Not Normal ' is classified as 0 as it could be case of other complication which is not phenumonia¶

In [23]:
pd.pivot_table(df_data,index=["class"], values=['patientId'], aggfunc='count')
Out[23]:
patientId
class
Lung Opacity 9555
No Lung Opacity / Not Normal 11821
Normal 8851
In [24]:
#sns.countplot(label_data["Target"].value_counts())
df_data["class"].value_counts().plot(kind='pie',autopct='%1.1f%%')
Out[24]:
<AxesSubplot: ylabel='class'>

From here we will now be creating a dataframe that will contain all the images and information about those X-Ray images¶

In [25]:
#directory information for the pneumonia and non-pneumonia cases 
parent_dir = 'D:/MyData/OneDrive/Study Materials/ML/Great Learning/Olympus/11 Capstone Project/1 Pneumonia Detection/data/'
stage_2_train_img_folder = "stage_2_train_images/"
stage_2_test_img_folder = "stage_2_test_images/"
saved_model_folder_path = 'models_bkp/'
test_image_item = '000db696-cf54-4385-b10b-6b16fbb3f985.dcm'
IMG_HEIGHT = 256
IMG_WIDTH = 256
In [26]:
# Creating a dataframe with the available columns
column_list = ["patientId", "Patient Sex", "Patient's Age", "View Position", "Image Size"]
train_file_meta_Data = pd.DataFrame(columns=column_list)
test_file_meta_Data = pd.DataFrame(columns=column_list)
In [27]:
print(f"Test image file name : \"{test_image_item}\"")
#displaying the metadata information that comes in a image file 
image_path = os.path.join(parent_dir,stage_2_train_img_folder,test_image_item) 
file = dicom.read_file(image_path)
file

Test image file name : "000db696-cf54-4385-b10b-6b16fbb3f985.dcm"
Out[27]:
Dataset.file_meta -------------------------------
(0002, 0000) File Meta Information Group Length  UL: 200
(0002, 0001) File Meta Information Version       OB: b'\x00\x01'
(0002, 0002) Media Storage SOP Class UID         UI: Secondary Capture Image Storage
(0002, 0003) Media Storage SOP Instance UID      UI: 1.2.276.0.7230010.3.1.4.8323329.4475.1517874307.936345
(0002, 0010) Transfer Syntax UID                 UI: JPEG Baseline (Process 1)
(0002, 0012) Implementation Class UID            UI: 1.2.276.0.7230010.3.0.3.6.0
(0002, 0013) Implementation Version Name         SH: 'OFFIS_DCMTK_360'
-------------------------------------------------
(0008, 0005) Specific Character Set              CS: 'ISO_IR 100'
(0008, 0016) SOP Class UID                       UI: Secondary Capture Image Storage
(0008, 0018) SOP Instance UID                    UI: 1.2.276.0.7230010.3.1.4.8323329.4475.1517874307.936345
(0008, 0020) Study Date                          DA: '19010101'
(0008, 0030) Study Time                          TM: '000000.00'
(0008, 0050) Accession Number                    SH: ''
(0008, 0060) Modality                            CS: 'CR'
(0008, 0064) Conversion Type                     CS: 'WSD'
(0008, 0090) Referring Physician's Name          PN: ''
(0008, 103e) Series Description                  LO: 'view: AP'
(0010, 0010) Patient's Name                      PN: '000db696-cf54-4385-b10b-6b16fbb3f985'
(0010, 0020) Patient ID                          LO: '000db696-cf54-4385-b10b-6b16fbb3f985'
(0010, 0030) Patient's Birth Date                DA: ''
(0010, 0040) Patient's Sex                       CS: 'F'
(0010, 1010) Patient's Age                       AS: '25'
(0018, 0015) Body Part Examined                  CS: 'CHEST'
(0018, 5101) View Position                       CS: 'AP'
(0020, 000d) Study Instance UID                  UI: 1.2.276.0.7230010.3.1.2.8323329.4475.1517874307.936344
(0020, 000e) Series Instance UID                 UI: 1.2.276.0.7230010.3.1.3.8323329.4475.1517874307.936343
(0020, 0010) Study ID                            SH: ''
(0020, 0011) Series Number                       IS: '1'
(0020, 0013) Instance Number                     IS: '1'
(0020, 0020) Patient Orientation                 CS: ''
(0028, 0002) Samples per Pixel                   US: 1
(0028, 0004) Photometric Interpretation          CS: 'MONOCHROME2'
(0028, 0010) Rows                                US: 1024
(0028, 0011) Columns                             US: 1024
(0028, 0030) Pixel Spacing                       DS: [0.168, 0.168]
(0028, 0100) Bits Allocated                      US: 8
(0028, 0101) Bits Stored                         US: 8
(0028, 0102) High Bit                            US: 7
(0028, 0103) Pixel Representation                US: 0
(0028, 2110) Lossy Image Compression             CS: '01'
(0028, 2114) Lossy Image Compression Method      CS: 'ISO_10918_1'
(7fe0, 0010) Pixel Data                          OB: Array of 100460 elements
In [28]:
# VIEWING XRAY IMAGE
ds = dicom.dcmread(image_path)
plt.imshow(ds.pixel_array)
Out[28]:
<matplotlib.image.AxesImage at 0x294c8fd3820>
In [29]:
#collecting dcm xray filenames in collection
train_images_dir = os.path.join(parent_dir,stage_2_train_img_folder)
test_images_dir = os.path.join(parent_dir,stage_2_test_img_folder)
train_images = [f for f in os.listdir(train_images_dir) if os.path.isfile(os.path.join(train_images_dir, f))]
test_images = [f for f in os.listdir(test_images_dir) if os.path.isfile(os.path.join(test_images_dir, f))]
print('5 Training images', train_images[:5])
print("=================================================================")
print('5 Test images', test_images[:5])

5 Training images ['0004cfab-14fd-4e49-80ba-63a80b6bddd6.dcm', '000924cf-0f8d-42bd-9158-1af53881a557.dcm', '000db696-cf54-4385-b10b-6b16fbb3f985.dcm', '000fe35a-2649-43d4-b027-e67796d412e0.dcm', '001031d9-f904-4a23-b3e5-2c088acd19c6.dcm']
=================================================================
5 Test images ['0000a175-0e68-4ca4-b1af-167204a7e0bc.dcm', '0005d3cc-3c3f-40b9-93c3-46231c3eb813.dcm', '000686d7-f4fc-448d-97a0-44fa9c5d3aa6.dcm', '000e3a7d-c0ca-4349-bb26-5af2d8993c3d.dcm', '00100a24-854d-423d-a092-edcf6179e061.dcm']
In [30]:
#this finction will add the metadata information to the dataframe and the we can analyze the information
def add_meta_data_to_df(df, loc, from_list):
    data = []
    for filename in from_list:
            imagePath = loc+filename
            data_row_img_data = dicom.read_file(imagePath)
            values = []
            values.append(data_row_img_data.PatientID)
            values.append(data_row_img_data.PatientSex)
            values.append(data_row_img_data.PatientAge)
            values.append(data_row_img_data.ViewPosition)
            values.append(f"{data_row_img_data.Rows}x{data_row_img_data.Columns}")
            zipped_val = dict(zip(column_list, values))
            df = df.append(zipped_val, True)
    return df
In [39]:
%%time
#loading image files from the train image folder location and collecting metadata patient age,sex,image size etc from there
train_file_meta_Data = add_meta_data_to_df(train_file_meta_Data, train_images_dir, train_images)
test_file_meta_Data = add_meta_data_to_df(test_file_meta_Data,test_images_dir,test_images)

CPU times: total: 2min 9s
Wall time: 4min 49s
In [40]:
print("train image data metadata information dataframe shape : ", train_file_meta_Data.shape)
print("test image data metadata information dataframe shape : ", test_file_meta_Data.shape)

train image data metadata information dataframe shape :  (26684, 5)
test image data metadata information dataframe shape :  (3000, 5)
In [41]:
train_file_meta_Data.head()
Out[41]:
patientId Patient Sex Patient's Age View Position Image Size
0 0004cfab-14fd-4e49-80ba-63a80b6bddd6 F 51 PA 1024x1024
1 000924cf-0f8d-42bd-9158-1af53881a557 F 19 AP 1024x1024
2 000db696-cf54-4385-b10b-6b16fbb3f985 F 25 AP 1024x1024
3 000fe35a-2649-43d4-b027-e67796d412e0 M 40 AP 1024x1024
4 001031d9-f904-4a23-b3e5-2c088acd19c6 M 57 PA 1024x1024
In [42]:
test_file_meta_Data.head()
Out[42]:
patientId Patient Sex Patient's Age View Position Image Size
0 0000a175-0e68-4ca4-b1af-167204a7e0bc F 46 PA 1024x1024
1 0005d3cc-3c3f-40b9-93c3-46231c3eb813 F 22 PA 1024x1024
2 000686d7-f4fc-448d-97a0-44fa9c5d3aa6 M 64 PA 1024x1024
3 000e3a7d-c0ca-4349-bb26-5af2d8993c3d F 75 PA 1024x1024
4 00100a24-854d-423d-a092-edcf6179e061 F 66 AP 1024x1024
In [43]:
train_file_meta_Data.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 26684 entries, 0 to 26683
Data columns (total 5 columns):
 #   Column         Non-Null Count  Dtype 
---  ------         --------------  ----- 
 0   patientId      26684 non-null  object
 1   Patient Sex    26684 non-null  object
 2   Patient's Age  26684 non-null  object
 3   View Position  26684 non-null  object
 4   Image Size     26684 non-null  object
dtypes: object(5)
memory usage: 1.0+ MB
In [44]:
#showing piechart for PA and AP data as piechart
train_file_meta_Data['View Position'].value_counts().plot(kind='pie',autopct='%1.1f%%')
Out[44]:
<AxesSubplot: ylabel='View Position'>
In [45]:
#showing piechart for PA and AP data as piechart
test_file_meta_Data['View Position'].value_counts().plot(kind='pie',autopct='%1.1f%%')
Out[45]:
<AxesSubplot: ylabel='View Position'>

test file and train file both show somewhat same proportion of view position¶

In [46]:
#Printing chart for patient's age
plt.title('Patient\'s Age')
sns.distplot(train_file_meta_Data["Patient's Age"],color='red')
Out[46]:
<AxesSubplot: title={'center': "Patient's Age"}, xlabel="Patient's Age", ylabel='Density'>
In [47]:
#Printing chart for patient's age
plt.title('Patient\'s Age')
sns.distplot(test_file_meta_Data["Patient's Age"],color='red')
Out[47]:
<AxesSubplot: title={'center': "Patient's Age"}, xlabel="Patient's Age", ylabel='Density'>
In [48]:
print('Number of train images:', len(train_images))
print('Number of test images:', len(test_images))

Number of train images: 26684
Number of test images: 3000
In [49]:
#method to show image with respective bounding box
def show_dicom_images_with_boxes(data):
    img_data = list(data.T.to_dict().values())
    f, ax = plt.subplots(3,3, figsize=(16,18))
    for i,data_row in enumerate(img_data):
        
        patientImage = data_row['patientId']+'.dcm'
        imagePath = f"{train_images_dir}/{patientImage}"
        
        drw_img_data = dicom.read_file(imagePath)
        drow_img = dicom.dcmread(imagePath)
        
        patid = data_row['patientId']
        trgt = data_row['Target']
        modality = drw_img_data.Modality
        age = drw_img_data.PatientAge
        sex = drw_img_data.PatientSex
        chrt_title = f'ID: {patid}\nModality: {modality} Age: {age} Sex: {sex} Target: {trgt}'
        
        ax[i//3, i%3].imshow(drow_img.pixel_array, cmap=plt.cm.bone) 
        ax[i//3, i%3].axis('off')
        ax[i//3, i%3].set_title(chrt_title)
        
        rows = df_data[df_data['patientId']==patid]
        box_data = list(rows.T.to_dict().values())
        
        for j, row in enumerate(box_data):
             ax[i//3, i%3].add_patch(Rectangle(xy=(row['x'], row['y']),
                         width=row['width'],height=row['height'], fc='none', lw=3,
                         color="red",alpha = 0.3))   
    plt.show()

Chest X-ray images of Normal cases¶

In [50]:
show_dicom_images_with_boxes(df_data[df_data['Target']==0].sample(9))

Chest X-ray images of Pneumonia cases¶

In [51]:
show_dicom_images_with_boxes(df_data[df_data['Target']==1].sample(9))

DCM FILE TO NUMPY CONVERSION¶

first creating train folder and creating 2 folders for Normal and Pneumonia cases¶

In [123]:
%%time
img_working_dirname = 'C:/Users/acer/ML_Projects/working_images'
img_working_dcm_to_jpg_train = img_working_dirname + '/pne_jpg_data/train/'
img_working_dcm_to_jpg_test = img_working_dirname + '/pne_jpg_data/test/'
img_working_split_ = img_working_dirname + '/d_/'
img_working_anno_dirname = img_working_dirname + '/annotation/'
img_working_csv_dirname = img_working_dirname + '/csv_files/'
#######FOR YOLO WORK############
img_working_yolo_dirname = img_working_dirname + '/yolo/'
img_working_yolo_images_dirname = img_working_dirname + '/yolo/images/'
img_working_yolo_labels_dirname = img_working_dirname + '/yolo/labels/'
img_working_yolo_images_train_dirname = img_working_dirname + '/yolo/images/train/'
img_working_yolo_labels_train_dirname = img_working_dirname + '/yolo/labels/train/'
img_working_yolo_images_test_dirname = img_working_dirname + '/yolo/images/test/'
img_working_yolo_labels_test_dirname = img_working_dirname + '/yolo/labels/test/'
img_working_yolo_images_val_dirname = img_working_dirname + '/yolo/images/val/'
img_working_yolo_labels_val_dirname = img_working_dirname + '/yolo/labels/val/'
#img_working_yolo_test_dirname = img_working_dirname + '/yolo/test/'

if not os.path.exists(img_working_dirname+"models_bkp"):
    os.mkdir(img_working_dirname+"models_bkp")  #empty test image folder

def create_image_directory_structure(rm_existing=False):
    if rm_existing:
        if os.path.exists("working_images"):
            shutil.rmtree("working_images") 
    if not os.path.exists("working_images"):
        os.mkdir("working_images")  #empty images folder 
    if not os.path.exists("working_images/pne_jpg_data"):
        os.mkdir("working_images/pne_jpg_data")  #empty images folder 
    if not os.path.exists("working_images/pne_jpg_data/train"):
        os.mkdir("working_images/pne_jpg_data/train")  #empty images folder 
    if not os.path.exists("working_images/pne_jpg_data/train/normal"):
        os.mkdir("working_images/pne_jpg_data/train/normal")  #empty images folder 
    if not os.path.exists("working_images/pne_jpg_data/train/pneumonia"):
        os.mkdir("working_images/pne_jpg_data/train/pneumonia")  #empty images folder 
    if not os.path.exists("working_images/pne_jpg_data/test"):
        os.mkdir("working_images/pne_jpg_data/test")  #empty images folder 
    
    ############# YOLO start
    #Yolo train Labels
    if not os.path.exists(img_working_yolo_dirname):
        os.mkdir(img_working_yolo_dirname)  #empty YOLO parent folder
    #Yolo images and lables parent    
    if not os.path.exists(img_working_yolo_images_dirname):
        os.mkdir(img_working_yolo_images_dirname)  #empty Yolo images folder
    if not os.path.exists(img_working_yolo_labels_dirname):
        os.mkdir(img_working_yolo_labels_dirname)  #empty Yolo Labels folder
        
    #Yolo train Labels and images
    if not os.path.exists(img_working_yolo_labels_train_dirname):
        os.mkdir(img_working_yolo_labels_train_dirname)  #empty labels train folder 
    if not os.path.exists(img_working_yolo_images_train_dirname):
        os.mkdir(img_working_yolo_images_train_dirname)  #empty yolo images Train folder 
    #Yolo test Labels and images 
    if not os.path.exists(img_working_yolo_images_test_dirname):
        os.mkdir(img_working_yolo_images_test_dirname)  #empty yolo images test folder 
    if not os.path.exists(img_working_yolo_labels_test_dirname):
        os.mkdir(img_working_yolo_labels_test_dirname)  #empty labels test folder 
    #Yolo val Labels and images 
    if not os.path.exists(img_working_yolo_images_val_dirname):
        os.mkdir(img_working_yolo_images_val_dirname)  #empty yolo images val folder 
    if not os.path.exists(img_working_yolo_labels_val_dirname):
        os.mkdir(img_working_yolo_labels_val_dirname)  #empty labels val folder 
        
    ############# YOLO END
    if not os.path.exists("working_images/annotation"):
        os.mkdir("working_images/annotation")  #empty images folder 

    if not os.path.exists("working_images/csv_files"):
        os.mkdir("working_images/csv_files")  #empty images folder 

    if not os.path.exists("working_images/d_"):#train image folder
        os.mkdir("working_images/d_") 
    if not os.path.exists("working_images/d_/train"):#train image folder
        os.mkdir("working_images/d_/train") 
    if not os.path.exists("working_images/d_/val"):#train image folder
        os.mkdir("working_images/d_/val")  
    if not os.path.exists("working_images/d_/test"):#train image folder
        os.mkdir("working_images/d_/test")  
    if not os.path.exists("working_images/d_/train/normal"):#image folder for normal cases  
        os.mkdir("working_images/d_/train/normal")
    if not os.path.exists("working_images/d_/train/pneumonia"):#image folder for pneumonia cases  
        os.mkdir("working_images/d_/train/pneumonia")  
    if not os.path.exists("working_images/d_/val/normal"):#image folder for normal cases  
        os.mkdir("working_images/d_/val/normal")
    if not os.path.exists("working_images/d_/val/pneumonia"):#image folder for pneumonia cases  
        os.mkdir("working_images/d_/val/pneumonia")  
    if not os.path.exists("working_images/d_/test/normal"):
        os.mkdir("working_images/d_/test/normal")  #empty test image folder
    if not os.path.exists("working_images/d_/test/pneumonia"):
        os.mkdir("working_images/d_/test/pneumonia")  #empty test image folder

CPU times: total: 0 ns
Wall time: 2 ms
In [53]:
#creating folder structure
create_image_directory_structure()
In [54]:
print(df_data.shape)
print(train_file_meta_Data.shape)

(30227, 7)
(26684, 5)
In [55]:
#file_meta_Data.rename(columns={"patientid":"patientId"},inplace=True)
df_data = df_data.join(train_file_meta_Data.set_index('patientId'),on="patientId",how="inner")
#df_data.patientId
In [56]:
print(df_data.shape)
df_data.head()
#file_meta_Data.head()

(30227, 11)
Out[56]:
patientId class x y width height Target Patient Sex Patient's Age View Position Image Size
0 0004cfab-14fd-4e49-80ba-63a80b6bddd6 No Lung Opacity / Not Normal NaN NaN NaN NaN 0 F 51 PA 1024x1024
1 00313ee0-9eaa-42f4-b0ab-c148ed3241cd No Lung Opacity / Not Normal NaN NaN NaN NaN 0 F 48 PA 1024x1024
2 00322d4d-1c29-4943-afc9-b6754be640eb No Lung Opacity / Not Normal NaN NaN NaN NaN 0 M 19 AP 1024x1024
3 003d8fa0-6bf1-40ed-b54c-ac657f8495c5 Normal NaN NaN NaN NaN 0 M 28 PA 1024x1024
4 00436515-870c-4b36-a041-de91049b9ab4 Lung Opacity 264.0 152.0 213.0 379.0 1 F 32 AP 1024x1024
In [57]:
#re-working on the data selection as now we want to consider both type of view positions
df_tmp_normal = df_data[df_data["Target"]==0]
df_tmp_pneumo = df_data[df_data["Target"]==1]
df_tmp_normal.reset_index(inplace=True)
df_tmp_pneumo.reset_index(inplace=True)

#sampling 5000 rows randomly from the collection from normal & penumonia infected case data
df_temp_normal_AP = df_tmp_normal[df_tmp_normal["View Position"] == "AP"]
df_temp_normal_AP.reset_index(inplace=True)
df_temp_normal_PA = df_tmp_normal[df_tmp_normal["View Position"] == "PA"]
df_temp_normal_PA.reset_index(inplace=True)

df_temp_pneumonia_AP = df_tmp_pneumo[df_tmp_pneumo["View Position"] == "AP"]
df_temp_pneumonia_AP.reset_index(inplace=True)
df_temp_pneumonia_PA = df_tmp_pneumo[df_tmp_pneumo["View Position"] == "PA"]
df_temp_pneumonia_PA.reset_index(inplace=True)
#print("Shape of all PA with normal cases",df_tmp_normal.shape)
#print("Shape of all PA with pneumonia cases",df_tmp_pneumo.shape)
print("Shape of all normal cases",df_tmp_normal.shape)
print("Shape of all pneumonia cases",df_tmp_pneumo.shape)

#for patid in file_meta_Data[file_meta_Data["View Position"]=="PA"]

Shape of all normal cases (20672, 12)
Shape of all pneumonia cases (9555, 12)

we have 1767 pneumonia cases and 13163 non penumonia cases (normal). We want to create a sample where we can have equal number of normal and pneumonia cases i.e. we may consider 1500 cases of pneumonia and 1500 case of normal. and because we have multiple entries for pneumonia cases for the same patient id beause of multiple annotation so the size may be more than 3000 of the final dataframe and for this reason we will select 1500 patient ids from normal and pneumonia cases from "file_meta_Data" dataframe¶

after re-considering all the cases whether the XRays are of PA or the AP type view position we found around 20672 case we have as normal cases and 9555 cases as pneumonia¶

In [58]:
df_tmp_pneumo["View Position"].value_counts()
Out[58]:
AP    7788
PA    1767
Name: View Position, dtype: int64
In [59]:
df_tmp_normal["View Position"].value_counts()
Out[59]:
PA    13163
AP     7509
Name: View Position, dtype: int64

After filtering Pneumonia cases and Normal cases we have 7788 AP view position data and 1767 PA cases we will create sample in such a way that when we model our data it will have equal number of PA/AP view position data in pneumonia and other one will be Normal cases.¶

we also observe that our normal cases also have "PA" view position as 13163 count and AP we have 7509 coumt. so our sample data from these normal case must have equal number of PA and AP view position data in our train same so that our model predict correctly.¶

If we choose a sample collection of 5000 image data than we may consider 2500 normal case images with 1250 patients with Xrays as PA and AP view position and 2500 pneumonia cases with 1250 PA view position and 1250 AP view Position¶

In [60]:
#Pneumonia cases only 
#df_temp_AP.join(,how="inner",on="patientId")
ii = 0
com_patid_ap = []
com_patid_ap_alrdy_exist = []
com_patid_pa = []
com_patid_pa_alrdy_exist = []
#common_patid_ap = []
for patid in list(df_temp_pneumonia_AP['patientId']):
    if patid not in com_patid_ap:
        com_patid_ap.append(patid)
    else:
        com_patid_ap_alrdy_exist.append(patid)
        
for patid in list(df_temp_pneumonia_PA['patientId']):
    if patid not in com_patid_pa:
        com_patid_pa.append(patid)
    else:
        com_patid_pa_alrdy_exist.append(patid)
        

print("Pneumonia \"PA\" View Position shape : ", df_temp_pneumonia_PA.shape," ~~~ distinct patientid count : ", len(com_patid_pa),"~~~ paientid found again : ",len(com_patid_ap_alrdy_exist))
print("Pneumonia \"AP\" View Position shape : ", df_temp_pneumonia_AP.shape," ~~~ distinct patientid count : ", len(com_patid_ap),"~~~ paientid found again : ",len(com_patid_pa_alrdy_exist))

Pneumonia "PA" View Position shape :  (1767, 13)  ~~~ distinct patientid count :  1348 ~~~ paientid found again :  3124
Pneumonia "AP" View Position shape :  (7788, 13)  ~~~ distinct patientid count :  4664 ~~~ paientid found again :  419
In [61]:
df_tmp_pneumo[df_tmp_pneumo["patientId"] == '00704310-78a8-4b38-8475-49f4573b2dbb']
Out[61]:
index patientId class x y width height Target Patient Sex Patient's Age View Position Image Size
2 10 00704310-78a8-4b38-8475-49f4573b2dbb Lung Opacity 323.0 577.0 160.0 104.0 1 M 75 PA 1024x1024
3 11 00704310-78a8-4b38-8475-49f4573b2dbb Lung Opacity 695.0 575.0 162.0 137.0 1 M 75 PA 1024x1024
In [62]:
#norma cases only 
#df_temp_AP.join(,how="inner",on="patientId")
ii = 0
com_patid_normal_ap = []
com_patid_normal_ap_alrdy_exist = []
com_patid_normal_pa = []
com_patid_normal_pa_alrdy_exist = []
#common_patid_ap = []
for patid in list(df_temp_normal_AP['patientId']):
    if patid not in com_patid_normal_ap:
        com_patid_normal_ap.append(patid)
    else:
        com_patid_normal_ap_alrdy_exist.append(patid)
        
for patid in list(df_temp_normal_PA['patientId']):
    if patid not in com_patid_normal_pa:
        com_patid_normal_pa.append(patid)
    else:
        com_patid_normal_pa_alrdy_exist.append(patid)
        

print("Normal \"PA\" View Position shape : ", df_temp_normal_PA.shape," ~~~ distinct patientid count : ", len(com_patid_normal_pa),"~~~ paientid found again : ",len(com_patid_normal_ap_alrdy_exist))
print("Normal \"AP\" View Position shape : ", df_temp_normal_AP.shape," ~~~ distinct patientid count : ", len(com_patid_normal_ap),"~~~ paientid found again : ",len(com_patid_normal_pa_alrdy_exist))

Normal "PA" View Position shape :  (13163, 13)  ~~~ distinct patientid count :  13163 ~~~ paientid found again :  0
Normal "AP" View Position shape :  (7509, 13)  ~~~ distinct patientid count :  7509 ~~~ paientid found again :  0

We found 1348 patients are pneumonia cases with XRay view position was "PA", 4664 patient counts are for "AP" View position so we will pick 1250 patientid from each PA and AP cases.¶

But in normal cases we 13163 patientid with PA view position and 7509 with AP view position. so in normal cases also we will pick 1250 normal cases of "PA" and 1250 of "AP" view position¶

In [63]:
%%time

#random pick from normal cases with AP position
ran_picked_indexes = random.sample([a for a in range(0,len(df_temp_normal_AP))], k=1400)
ran_picked_normal_ap_patid = df_temp_normal_AP.loc[ran_picked_indexes,["patientId"]].values
#random pick from normal cases with PA position
ran_picked_indexes = random.sample([a for a in range(0,len(df_temp_normal_PA))], k=1400)
ran_picked_normal_pa_patid = df_temp_normal_PA.loc[ran_picked_indexes,["patientId"]].values
#random pick from pneumonia cases with AP position
ran_picked_indexes = random.sample([a for a in range(0,len(df_temp_pneumonia_AP))], k=1400)
ran_picked_pneumo_ap_patid = df_temp_pneumonia_AP.loc[ran_picked_indexes,["patientId"]].values
#random pick from pneumonia cases with PA position
ran_picked_indexes = random.sample([a for a in range(0,len(df_temp_pneumonia_PA))], k=1348)
ran_picked_pneumo_pa_patid = df_temp_pneumonia_PA.loc[ran_picked_indexes,["patientId"]].values


print("Normal AP shape : ",ran_picked_normal_ap_patid.shape)
print("Normal PA shape : ",ran_picked_normal_pa_patid.shape)
print("Pneumonia AP shape : ",ran_picked_pneumo_ap_patid.shape)
print("Pneumonia PA shape : ",ran_picked_pneumo_pa_patid.shape)

# will now transfer dcm files from parent location to the respective train folders for normal and pneumonia cases
df_data_ap_trg0 = df_temp_normal_AP.loc[df_temp_normal_AP.apply(lambda x: x.patientId in ran_picked_normal_ap_patid, axis=1)]
df_data_pa_trg0 = df_temp_normal_PA.loc[df_temp_normal_PA.apply(lambda x: x.patientId in ran_picked_normal_pa_patid, axis=1)]
df_data_pa_trg1 = df_temp_pneumonia_PA.loc[df_temp_pneumonia_PA.apply(lambda x: x.patientId in ran_picked_pneumo_pa_patid, axis=1)]
df_data_ap_trg1 = df_temp_pneumonia_AP.loc[df_temp_pneumonia_AP.apply(lambda x: x.patientId in ran_picked_pneumo_ap_patid, axis=1)]
print("final sample AP normal cases: ", df_data_ap_trg0.shape)
print("final sample PA normal cases: ", df_data_pa_trg0.shape)
print("final sample PA penumo cases: ", df_data_pa_trg1.shape)
print("final sample AP penumo cases: ", df_data_ap_trg1.shape)
df_combined = [df_data_ap_trg0,df_data_pa_trg0,df_data_pa_trg1,df_data_ap_trg1]
df_data_final = pd.concat(df_combined)
print("final sample combined: ", df_data_final.shape)
#shuffling the dataframe rows
df_data_final = df_data_final.sample(frac = 1)
df_data_final.set_index("index",inplace=True)
df_data_final.reset_index(inplace=True,drop=True)
df_data_final.head()

Normal AP shape :  (1400, 1)
Normal PA shape :  (1400, 1)
Pneumonia AP shape :  (1400, 1)
Pneumonia PA shape :  (1348, 1)
final sample AP normal cases:  (1400, 13)
final sample PA normal cases:  (1400, 13)
final sample PA penumo cases:  (1519, 13)
final sample AP penumo cases:  (2375, 13)
final sample combined:  (6694, 13)
CPU times: total: 828 ms
Wall time: 809 ms
Out[63]:
level_0 patientId class x y width height Target Patient Sex Patient's Age View Position Image Size
0 3184 48a44f52-7fb3-4384-a76b-f6170b03a357 Normal NaN NaN NaN NaN 0 F 50 AP 1024x1024
1 2509 49b95513-daab-49bb-bc6e-c5254ab1bc07 Lung Opacity 599.0 202.0 216.0 537.0 1 F 67 AP 1024x1024
2 12362 a985f470-12d9-48b0-bb1a-5bdffc657c47 No Lung Opacity / Not Normal NaN NaN NaN NaN 0 F 58 PA 1024x1024
3 7670 75ac9b0b-6f99-49f1-865c-3ccf54144b1a No Lung Opacity / Not Normal NaN NaN NaN NaN 0 M 65 AP 1024x1024
4 1239 33c66fa9-6a08-4d0d-b309-88ce91a217c9 Lung Opacity 193.0 468.0 209.0 120.0 1 F 53 PA 1024x1024
In [64]:
df_data_final["patientId"].nunique()
Out[64]:
5219

Checking the data distribution in our sample data¶

Patient Gender¶

In [65]:
df_data_final["Patient Sex"].value_counts().plot(kind='pie',autopct='%1.1f%%')
Out[65]:
<AxesSubplot: ylabel='Patient Sex'>

Patient's Age distribution¶

In [66]:
#Printing chart for patient's age
plt.title('Patient\'s Age')
sns.distplot(df_data_final["Patient's Age"],color='red')
Out[66]:
<AxesSubplot: title={'center': "Patient's Age"}, xlabel="Patient's Age", ylabel='Density'>
In [67]:
print(train_images_dir)
print(test_images_dir)

D:/MyData/OneDrive/Study Materials/ML/Great Learning/Olympus/11 Capstone Project/1 Pneumonia Detection/data/stage_2_train_images/
D:/MyData/OneDrive/Study Materials/ML/Great Learning/Olympus/11 Capstone Project/1 Pneumonia Detection/data/stage_2_test_images/
In [68]:
def get_sharpened_and_resized_array_of_image(img_arr_):
    #https://stackoverflow.com/a/54465855/1225413
    #https://stackoverflow.com/a/6917185/1225413
    """
    Since the given dataset images are grayscale, we can improve the performance by sharpening. Also, we can reduce the 
    image size from 1024 *1024 to 256*256, cause their is rarely any improvement shown in model performance.
    """
    kernel = np.array([[-1,-1,-1], 
                       [-1, 9,-1],
                       [-1,-1,-1]])
    #resized array
    resized_img_ = cv2.resize(img_arr_, (IMG_HEIGHT, IMG_WIDTH), interpolation=cv2.INTER_AREA)

    #appyling kernel to sharpen the image
    sharpened_ = cv2.filter2D(resized_img_, -1, kernel)
    sharpened_ =  (255.0 / sharpened_.max() * (sharpened_ - sharpened_.min())).astype(np.uint8)
    return sharpened_
In [69]:
df_data_final.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 6694 entries, 0 to 6693
Data columns (total 12 columns):
 #   Column         Non-Null Count  Dtype  
---  ------         --------------  -----  
 0   level_0        6694 non-null   int64  
 1   patientId      6694 non-null   object 
 2   class          6694 non-null   object 
 3   x              3894 non-null   float64
 4   y              3894 non-null   float64
 5   width          3894 non-null   float64
 6   height         3894 non-null   float64
 7   Target         6694 non-null   int64  
 8   Patient Sex    6694 non-null   object 
 9   Patient's Age  6694 non-null   object 
 10  View Position  6694 non-null   object 
 11  Image Size     6694 non-null   object 
dtypes: float64(4), int64(2), object(6)
memory usage: 627.7+ KB
In [70]:
#creating folder structure
create_image_directory_structure(True)
In [71]:
pneumonia_case_images = []
normal_case_images = []
#for now we will be using only 5000+ images for normal and pneumonia cases randomly  
def transfer_dcm_to_jpg_wd_folder(image_type):
    images_path = os.listdir(train_images_dir)
    non_found_images = []

    ii = 0
    
    for patid in list(df_data_final["patientId"].unique()):
        
        xrdcm_fname = patid + '.dcm' #dcm filename
        xrjpg_fname = patid + '.' + image_type 
        dcm_img_path = os.path.join(train_images_dir, xrdcm_fname)
        copyto_loc = ''
        df_data_final.loc[df_data_final["patientId"]==patid]
        trgt=0
        for i,row in df_data_final[df_data_final["patientId"]==patid].iterrows():
              trgt = row["Target"]     
        if os.path.exists(dcm_img_path):     
            if trgt == 1:
                copyto_loc = os.path.join(img_working_dcm_to_jpg_train+"/pneumonia/",xrjpg_fname)
                pneumonia_case_images.append(xrjpg_fname)
            else:
                copyto_loc = os.path.join(img_working_dcm_to_jpg_train+"/normal/",xrjpg_fname)  
                normal_case_images.append(xrjpg_fname)

            ds = dicom.dcmread(dcm_img_path)
            arr_img_ = ds.pixel_array
            cv2.imwrite(copyto_loc, arr_img_)
        else:
            non_found_images.append(xrdcm_fname)
        ii += 1
#         if ii > 100:
#             break
    
    print("total converted to png from dcm :",ii)
    print("total normal cases converted to png from dcm :",len(normal_case_images))
    print("total pneumonia cases converted to png from dcm :",len(pneumonia_case_images))
    print("total not found images :",len(non_found_images))
#transfer_dcm_to_jpg_wd_folder('jpg')
In [72]:
%time
#converting and transfering images to train folder
transfer_dcm_to_jpg_wd_folder('jpg')

CPU times: total: 0 ns
Wall time: 0 ns
total converted to png from dcm : 5219
total normal cases converted to png from dcm : 2800
total pneumonia cases converted to png from dcm : 2419
total not found images : 0
In [73]:
def traindata_to_trainvaltest_split():

    val_ratio = 0.20
    test_ratio = 0.20

    for cls in ['normal','pneumonia']:
        # Creating partitions of the data after shuffeling
        src = img_working_dcm_to_jpg_train +"//" + cls  # Folder to copy images from

        allFileNames = os.listdir(src)
        np.random.shuffle(allFileNames)
        train_imgfiles, val_imgfiles, test_imgfiles = np.split(np.array(allFileNames),
                                                                  [int(len(allFileNames) * (1 - (val_ratio + test_ratio))),
                                                                   int(len(allFileNames) * (1 - val_ratio)),
                                                                   ])
        
        train_imgfiles = [src + '//' + name for name in train_imgfiles.tolist()]
        val_imgfiles = [src + '//' + name for name in val_imgfiles.tolist()]
        test_imgfiles = [src + '//' + name for name in test_imgfiles.tolist()]
        
        print(f'{cls} Total images: '+ str(len(allFileNames)))
        print(f'{cls} Training    : '+ str(len(train_imgfiles)))
        print(f'{cls} Validation  : '+  str(len(val_imgfiles)))
        print(f'{cls} Testing     : '+ str(len(test_imgfiles)))
        
        for name in train_imgfiles:
            shutil.copy(name, img_working_split_ + '/train//' + cls)

        for name in val_imgfiles:
            shutil.copy(name, img_working_split_ + '/val//' + cls)

        for name in test_imgfiles:
            shutil.copy(name, img_working_split_ + '/test//' + cls)

traindata_to_trainvaltest_split()

normal Total images: 2800
normal Training    : 1680
normal Validation  : 560
normal Testing     : 560
pneumonia Total images: 2419
pneumonia Training    : 1451
pneumonia Validation  : 484
pneumonia Testing     : 484
In [74]:
#for now we will be using only 2000 images for normal cases and 2000 pneumonia cases randomly  
def create_dcm_to_train_folder(image_type):
    images_path = os.listdir(train_images_dir)
    normal_count = 0
    pneumo_count = 0
    non_found_images = []
    
    for n, img_fname in enumerate(images_path):
        img_patid = img_fname.replace('.dcm','')
        img_fname_ = img_fname.replace('.dcm','.' + image_type)
        dcm_image_loc =  train_images_dir
        copy_filepath = ""
        ds = dicom.dcmread(os.path.join(train_images_dir, img_fname))
        
        # Normal PA
        if img_patid in ran_picked_normal_ap_patid:
            if os.path.exists(dcm_image_loc):
                #get the image resized and apply sharpening kernel
                arr_img_ = ds.pixel_array #get_sharpened_and_resized_array_of_image(ds.pixel_array) 
                cv2.imwrite(os.path.join(img_working_dirname+"/train/normal/",img_fname_), arr_img_)
                normal_count = normal_count + 1
            else:
                non_found_images.append("normal_"+dcm_image_loc)
            #print('normal')

        
        if img_patid in ran_picked_normal_pa_patid:
            if os.path.exists(dcm_image_loc):
                #get the image resized and apply sharpening kernel
                arr_img_ = ds.pixel_array # get_sharpened_and_resized_array_of_image(ds.pixel_array)
                cv2.imwrite(os.path.join(img_working_dirname+"/train/normal/",img_fname_), arr_img_)
                #cv2.imwrite(copyto_filepath, arr_img_)
                normal_count = normal_count + 1
            else:
                non_found_images.append("normal_"+dcm_image_loc)
            #print('normal')

        if img_patid in ran_picked_pneumo_ap_patid:
            if os.path.exists(dcm_image_loc):
                #get the image resized and apply sharpening kernel
                arr_img_ = ds.pixel_array #get_sharpened_and_resized_array_of_image(ds.pixel_array)
                cv2.imwrite(os.path.join(img_working_dirname+"/train/pneumonia/",img_fname_), arr_img_)
                #cv2.imwrite(copyto_filepath, arr_img_)
                pneumo_count = pneumo_count + 1
            else:
                non_found_images.append("pneumo_"+dcm_image_loc)
                
        if img_patid in ran_picked_pneumo_pa_patid:
            if os.path.exists(dcm_image_loc):
                #get the image resized and apply sharpening kernel
                arr_img_ = ds.pixel_array #get_sharpened_and_resized_array_of_image(ds.pixel_array)
                cv2.imwrite(os.path.join(img_working_dirname+"/train/pneumonia/",img_fname_), arr_img_)
                #cv2.imwrite(copyto_filepath, arr_img_)
                pneumo_count = pneumo_count + 1
            else:
                non_found_images.append("pneumo_"+dcm_image_loc)
            #print('pneumo')
    #     if ii > 20:
    #         break  
    print("total normal cases converted to png from dcm :",normal_count)
    print("total pneumo cases converted to png from dcm :",pneumo_count)
    print("total notfound images :",len(non_found_images))
In [75]:
test_images_dir
Out[75]:
'D:/MyData/OneDrive/Study Materials/ML/Great Learning/Olympus/11 Capstone Project/1 Pneumonia Detection/data/stage_2_test_images/'
In [76]:
#test images to working directory with resized format 
def create_dcm_to_test_folder(image_type):
    ii = 0
    if os.path.exists(img_working_dcm_to_jpg_test) and os.path.exists(test_images_dir):
        images_path = os.listdir(test_images_dir)
        for n, img_fname in enumerate(images_path): 
            img_patid = img_fname.replace('.dcm','')
            img_fname_jpg = img_fname.replace('.dcm','.' + image_type)
            img_path = test_images_dir+"/"+ img_fname
            img_ = dicom.read_file(img_path)
            arr_img_ = img_.pixel_array
            #get the image resized and apply sharpening kernel
            #arrimg_ = #get_sharpened_and_resized_array_of_image(arr_img_)
            cv2.imwrite(os.path.join(img_working_dcm_to_jpg_test,img_fname_jpg), arr_img_)
            #print(os.path.join(img_working_dcm_to_jpg_test,img_fname_jpg))
            ii = ii + 1
#             if ii > 2:
#                 break
        #print(ii, f" dcm files converted and resized({IMG_HEIGHT} X {IMG_WIDTH}) to png at :",copyto_loc_)
        print(ii, f" dcm files converted to png at :",img_working_dcm_to_jpg_test)
    else:
        print("locations not found.")
In [77]:
%%time
create_dcm_to_test_folder('jpg')

3000  dcm files converted to png at : C:/Users/acer/ML_Projects/working_images/pne_jpg_data/test/
CPU times: total: 43 s
Wall time: 43.2 s

Creating Data Generators¶

In [78]:
def get_random_test_image_from_test_folder(_test_image_class,n_count):
    fld_path = img_working_split_ + 'test/' + _test_image_class
    #print(fld_path)
    _image_names_collection = []
    if os.path.exists(fld_path):
        images_path = os.listdir(fld_path)
        ran_picked_image_indices = random.sample([a for a in range(0,len(images_path))], k=n_count)
        for n, img_fname in enumerate(images_path):
            if n in ran_picked_image_indices:
                f = fld_path + "/" + img_fname
                _image_names_collection.append(f)
    else :
        print('no hello')
    return _image_names_collection
        
#print(img_working_dcm_to_jpg_test)
#get_random_test_image_from_test_folder('pneumonia',10)
#get_random_test_image_from_test_folder('normal',10)
In [79]:
#path for the image to see if it predicts correct class
_test_images_normal = get_random_test_image_from_test_folder('normal',6)
_test_images_pneumo = get_random_test_image_from_test_folder('pneumonia',6)
In [80]:
print(img_working_split_)
# Create data generator for training data with data augmentation and normalizing all values by 255
train_image_datagen = ImageDataGenerator(rescale = 1./255,
                                   rotation_range=30,
                                   shear_range = 0.05,
                                   zoom_range = 0.3,
                                   brightness_range=(1.2,1.5),
                                   horizontal_flip = True)

train_generator = train_image_datagen.flow_from_directory(img_working_split_+"/train",
                                                 target_size = (IMG_HEIGHT, IMG_WIDTH),
                                                 batch_size = 16,
                                                 class_mode = 'binary',
                                                 shuffle=True)


val_image_datagen = ImageDataGenerator(rescale = 1./255)
val_generator = val_image_datagen.flow_from_directory(img_working_split_+"/val",
                                                 target_size = (IMG_HEIGHT, IMG_WIDTH),
                                                 batch_size = 16,
                                                 class_mode = 'binary',
                                                 shuffle=True)

# # Setting testing data generator's source directory
test_image_datagen = ImageDataGenerator(rescale = 1./255)
test_generator = test_image_datagen.flow_from_directory(img_working_split_+"/test",
                                            target_size = (IMG_HEIGHT, IMG_WIDTH),
                                            batch_size = 16,
                                            class_mode = 'binary')
############################################## VGG16 ######################################
train_image_datagen_vgg16 = ImageDataGenerator(rescale = 1./255,
                                   #rotation_range=30,
                                   #shear_range = 0.05,
                                   zoom_range = 0.3,
                                   brightness_range=(1.2,1.5),
                                   horizontal_flip = True)


train_generator_vgg16 = train_image_datagen_vgg16.flow_from_directory(img_working_split_+"/train",
                                                 target_size = (350,350), #(IMG_HEIGHT, IMG_WIDTH),
                                                 batch_size = 16,
                                                 class_mode = 'binary',
                                                 shuffle=True)

val_generator_vgg16 = val_image_datagen.flow_from_directory(img_working_split_+"/val",
                                                 target_size = (350,350),
                                                 batch_size = 16,
                                                 class_mode = 'binary',
                                                 shuffle=True)

test_generator_vgg16 = test_image_datagen.flow_from_directory(img_working_split_+"/test",
                                            target_size = (350,350),
                                            batch_size = 16,
                                            class_mode = 'binary')

C:/Users/acer/ML_Projects/working_images/d_/
Found 3131 images belonging to 2 classes.
Found 1044 images belonging to 2 classes.
Found 1044 images belonging to 2 classes.
Found 3131 images belonging to 2 classes.
Found 1044 images belonging to 2 classes.
Found 1044 images belonging to 2 classes.
In [81]:
## Accuracy and Loss plots
def show_accuracy_charts(hist):
    acc_ = hist.history['accuracy']
    val_acc_ = hist.history['val_accuracy']
    loss_ = hist.history['loss']
    val_loss_ = hist.history['val_loss']

    epochs   = range(len(acc_)) # Get number of epochs

    plt.plot  (epochs, acc_, label = 'training accuracy')
    plt.plot  (epochs, val_acc_, label = 'validation accuracy')
    plt.title ('Training and validation accuracy')
    plt.legend(loc = 'lower right')
    plt.figure()

    plt.plot  (epochs, loss_, label = 'training loss')
    plt.plot  (epochs, val_loss_, label = 'validation loss')
    plt.legend(loc = 'upper right')

Building a Model¶

Test Model1¶

In [82]:
#Test Model1
model1 = Sequential()
model1.add(Conv2D(filters = 32,kernel_size=(3,3),activation = 'relu' , input_shape = (IMG_HEIGHT,IMG_WIDTH,3)))

model1.add(Conv2D(filters = 64,kernel_size=(3,3),activation = 'relu'))
model1.add(MaxPool2D(pool_size=(2,2)))
model1.add(Dropout(0.25))

model1.add(Conv2D(filters = 64,kernel_size=(3,3),activation = 'relu'))
model1.add(MaxPool2D(pool_size=(2,2)))
model1.add(Dropout(0.25))

model1.add(Conv2D(filters = 128,kernel_size=(3,3),activation = 'relu'))
model1.add(MaxPool2D(pool_size=(2,2)))
model1.add(MaxPool2D(pool_size=(2,2)))
model1.add(Dropout(0.25))

model1.add(Flatten())
model1.add(Dense(units=64,activation='relu'))
model1.add(Dropout(0.5))
model1.add(Dense(units = 1 , activation = 'sigmoid'))
In [201]:
optimizer = Adam()
model1.compile(optimizer='adam', loss = 'binary_crossentropy', metrics = ['accuracy'])
model1.summary()

Model: "sequential_2"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d_6 (Conv2D)           (None, 254, 254, 32)      896       
                                                                 
 conv2d_7 (Conv2D)           (None, 252, 252, 64)      18496     
                                                                 
 max_pooling2d_4 (MaxPooling  (None, 126, 126, 64)     0         
 2D)                                                             
                                                                 
 dropout (Dropout)           (None, 126, 126, 64)      0         
                                                                 
 conv2d_8 (Conv2D)           (None, 124, 124, 64)      36928     
                                                                 
 max_pooling2d_5 (MaxPooling  (None, 62, 62, 64)       0         
 2D)                                                             
                                                                 
 dropout_1 (Dropout)         (None, 62, 62, 64)        0         
                                                                 
 conv2d_9 (Conv2D)           (None, 60, 60, 128)       73856     
                                                                 
 max_pooling2d_6 (MaxPooling  (None, 30, 30, 128)      0         
 2D)                                                             
                                                                 
 max_pooling2d_7 (MaxPooling  (None, 15, 15, 128)      0         
 2D)                                                             
                                                                 
 dropout_2 (Dropout)         (None, 15, 15, 128)       0         
                                                                 
 flatten_1 (Flatten)         (None, 28800)             0         
                                                                 
 dense_2 (Dense)             (None, 64)                1843264   
                                                                 
 dropout_3 (Dropout)         (None, 64)                0         
                                                                 
 dense_3 (Dense)             (None, 1)                 65        
                                                                 
=================================================================
Total params: 1,973,505
Trainable params: 1,973,505
Non-trainable params: 0
_________________________________________________________________
In [200]:
model_path1 = os.path.join(saved_model_folder_path,"best_model1.h5")
print(model_path1)
early_stop1 = tf.keras.callbacks.EarlyStopping(monitor="val_accuracy",min_delta=0.01,patience=10,verbose=1,mode='auto')
checkpoint1 = ModelCheckpoint(model_path1,monitor='val_accuracy',save_best_only=True, mode='auto',verbose=1)
reduce_lr1 = ReduceLROnPlateau(monitor='accuracy',factor=0.1,patience=5,min_lr=0.00001,model='auto')
callbacks1 = [early_stop1,checkpoint1,reduce_lr1]

models_bkp/best_model1.h5
In [202]:
%%time
history1 = model1.fit_generator(generator=train_generator,epochs=30,validation_data=val_generator,validation_steps=16,workers=15,callbacks=callbacks1)

Epoch 1/30
195/195 [==============================] - ETA: 0s - loss: 0.7115 - accuracy: 0.5377
Epoch 1: val_accuracy improved from -inf to 0.59375, saving model to models_bkp\best_model1.h5
195/195 [==============================] - 26s 115ms/step - loss: 0.7115 - accuracy: 0.5377 - val_loss: 0.6863 - val_accuracy: 0.5938 - lr: 0.0010
Epoch 2/30
195/195 [==============================] - ETA: 0s - loss: 0.6918 - accuracy: 0.5342
Epoch 2: val_accuracy did not improve from 0.59375
195/195 [==============================] - 22s 108ms/step - loss: 0.6918 - accuracy: 0.5342 - val_loss: 0.6923 - val_accuracy: 0.5195 - lr: 0.0010
Epoch 3/30
195/195 [==============================] - ETA: 0s - loss: 0.6902 - accuracy: 0.5490
Epoch 3: val_accuracy did not improve from 0.59375
195/195 [==============================] - 21s 106ms/step - loss: 0.6902 - accuracy: 0.5490 - val_loss: 0.6911 - val_accuracy: 0.5312 - lr: 0.0010
Epoch 4/30
195/195 [==============================] - ETA: 0s - loss: 0.6849 - accuracy: 0.5528
Epoch 4: val_accuracy did not improve from 0.59375
195/195 [==============================] - 22s 106ms/step - loss: 0.6849 - accuracy: 0.5528 - val_loss: 0.6885 - val_accuracy: 0.5273 - lr: 0.0010
Epoch 5/30
195/195 [==============================] - ETA: 0s - loss: 0.6773 - accuracy: 0.5737
Epoch 5: val_accuracy improved from 0.59375 to 0.62500, saving model to models_bkp\best_model1.h5
195/195 [==============================] - 21s 104ms/step - loss: 0.6773 - accuracy: 0.5737 - val_loss: 0.6534 - val_accuracy: 0.6250 - lr: 0.0010
Epoch 6/30
195/195 [==============================] - ETA: 0s - loss: 0.6742 - accuracy: 0.5737
Epoch 6: val_accuracy did not improve from 0.62500
195/195 [==============================] - 20s 100ms/step - loss: 0.6742 - accuracy: 0.5737 - val_loss: 0.6714 - val_accuracy: 0.5625 - lr: 0.0010
Epoch 7/30
195/195 [==============================] - ETA: 0s - loss: 0.6764 - accuracy: 0.5776
Epoch 7: val_accuracy did not improve from 0.62500
195/195 [==============================] - 20s 100ms/step - loss: 0.6764 - accuracy: 0.5776 - val_loss: 0.6810 - val_accuracy: 0.5469 - lr: 0.0010
Epoch 8/30
195/195 [==============================] - ETA: 0s - loss: 0.6810 - accuracy: 0.5776
Epoch 8: val_accuracy did not improve from 0.62500
195/195 [==============================] - 20s 99ms/step - loss: 0.6810 - accuracy: 0.5776 - val_loss: 0.6866 - val_accuracy: 0.5391 - lr: 0.0010
Epoch 9/30
195/195 [==============================] - ETA: 0s - loss: 0.6732 - accuracy: 0.5856
Epoch 9: val_accuracy did not improve from 0.62500
195/195 [==============================] - 20s 99ms/step - loss: 0.6732 - accuracy: 0.5856 - val_loss: 0.6818 - val_accuracy: 0.5469 - lr: 0.0010
Epoch 10/30
195/195 [==============================] - ETA: 0s - loss: 0.6647 - accuracy: 0.5895
Epoch 10: val_accuracy did not improve from 0.62500
195/195 [==============================] - 20s 101ms/step - loss: 0.6647 - accuracy: 0.5895 - val_loss: 0.6595 - val_accuracy: 0.5859 - lr: 0.0010
Epoch 11/30
195/195 [==============================] - ETA: 0s - loss: 0.6619 - accuracy: 0.5978
Epoch 11: val_accuracy did not improve from 0.62500
195/195 [==============================] - 20s 100ms/step - loss: 0.6619 - accuracy: 0.5978 - val_loss: 0.6538 - val_accuracy: 0.6055 - lr: 0.0010
Epoch 12/30
195/195 [==============================] - ETA: 0s - loss: 0.6621 - accuracy: 0.5965
Epoch 12: val_accuracy did not improve from 0.62500
195/195 [==============================] - 20s 101ms/step - loss: 0.6621 - accuracy: 0.5965 - val_loss: 0.6696 - val_accuracy: 0.5820 - lr: 0.0010
Epoch 13/30
195/195 [==============================] - ETA: 0s - loss: 0.6567 - accuracy: 0.6039
Epoch 13: val_accuracy did not improve from 0.62500
195/195 [==============================] - 20s 99ms/step - loss: 0.6567 - accuracy: 0.6039 - val_loss: 0.6585 - val_accuracy: 0.5898 - lr: 0.0010
Epoch 14/30
195/195 [==============================] - ETA: 0s - loss: 0.6550 - accuracy: 0.6113
Epoch 14: val_accuracy improved from 0.62500 to 0.63281, saving model to models_bkp\best_model1.h5
195/195 [==============================] - 20s 100ms/step - loss: 0.6550 - accuracy: 0.6113 - val_loss: 0.6447 - val_accuracy: 0.6328 - lr: 0.0010
Epoch 15/30
195/195 [==============================] - ETA: 0s - loss: 0.6555 - accuracy: 0.6100
Epoch 15: val_accuracy did not improve from 0.63281
195/195 [==============================] - 20s 100ms/step - loss: 0.6555 - accuracy: 0.6100 - val_loss: 0.6465 - val_accuracy: 0.6055 - lr: 0.0010
Epoch 15: early stopping
CPU times: total: 22min 26s
Wall time: 5min 14s
In [203]:
def print_performance_results(model_):
    
    true_lbls = test_generator.labels
    pred_lbls = np.squeeze(np.array(model_.predict_generator(test_generator) >= .5,dtype=np.int16))
    print()
    print("Classes: ",test_generator.class_indices)
    print()
    print("=================================================================")
    cm_ = confusion_matrix(true_lbls,pred_lbls)
    plt.figure(figsize=(4,4))
    
    sns.heatmap(cm_,annot=True,fmt='g',vmin=0,cmap='mako',cbar=False)
    plt.xlabel("Predicted")
    plt.ylabel("Actual")
    plt.xticks(np.arange(2)+0.5,['Normal','Pneumonia'],fontsize=10)
    plt.yticks(np.arange(2)+0.5,['Normal','Pneumonia'],fontsize=10)
    plt.show()
    
    res = model_.evaluate_generator(test_generator,verbose=0)
    acc = res[0]
    auc = res[1]
    tn,fp,fn,tp = cm_.ravel()
    precision = tp/(tp+fp) #whatever we predicted how many of them were correct
    recall = tp/(tp+fn) #how many of the true positive werepredicted correctly
    print("=================================================================")
    print("Accuracy:      {:.2f}%".format(acc*100))
    print("AUC:           {:.2f}%".format(auc*100))
    print("Precision:     {:.2f}%".format(precision*100))
    print("Recall:        {:.2f}%".format(recall*100))
    print("=================================================================")
In [325]:
#im_ = cv2.imread('C:/Users/acer/ML_Projects/working_images/d_/test/normal/037e120a-24fa-4f9c-8813-111136e3c288.jpg')
def predict_result(img_,model_,confi_):
    image_ = cv2.cvtColor(img_,cv2.COLOR_BGR2RGB)
    #because our model expects batches, so creating a batch of 1 
    image_ = cv2.resize(image_,(IMG_HEIGHT,IMG_WIDTH))
    image_ar = tf.keras.utils.img_to_array(image_)/255
    #print(image_ar.shape)
    image_ar = np.expand_dims(image_ar,axis=0)
    pred = model_.predict(image_ar)
    print(pred)
    _res = ''
    if pred > confi_:
        _res = "Pneumonia"
    else:
        _res = "Normal"
    return _res,pred
#np.set_printoptions(suppress=True)
#result
#predict_result(im_,model1)
In [348]:
_CONFIDENCE_CRITERIA = .55
#method to show image with respective bounding box
def show_test_images_pred_img_panel(model_):
    _images_row = 4
    _images_col = 3
    f, ax = plt.subplots(_images_row,_images_col, figsize=(12,12))
    ic_ = 0
    for i,data_row in enumerate(_test_images_normal):
        image_ = cv2.imread(data_row)
        res,_pred = predict_result(image_,model_,_CONFIDENCE_CRITERIA)
        title_ = f"{ic_+1}- A: Normal P: " + res + " %.2f%%"%(_pred*100)
        ax[ic_//_images_col, ic_%_images_col].imshow(image_, cmap=plt.cm.bone) 
        ax[ic_//_images_col, ic_%_images_col].axis('off')
        ax[ic_//_images_col, ic_%_images_col].set_title(title_, fontsize=8)
        ic_ += 1
        
    for i,data_row in enumerate(_test_images_pneumo):
        image_ = cv2.imread(data_row)
        res,_pred = predict_result(image_,model_,_CONFIDENCE_CRITERIA)
        title_ = f"{ic_+1}- A: Pneumonia P: " + res + " %.2f%%"%(_pred*100)
        ax[ic_//_images_col, ic_%_images_col].imshow(image_, cmap=plt.cm.bone) 
        ax[ic_//_images_col, ic_%_images_col].axis('off')
        ax[ic_//_images_col, ic_%_images_col].set_title(title_, fontsize=8)
        ic_ += 1
        
    """
        for j, row in enumerate(box_data):
             ax[i//3, i%3].add_patch(Rectangle(xy=(row['x'], row['y']),
                         width=row['width'],height=row['height'], fc='none', lw=3,
                         color="red",alpha = 0.3))  
    """
    plt.show()
    
#_test_images_normal = get_random_test_image_from_test_folder('normal',4)
#_test_images_pneumo = get_random_test_image_from_test_folder('pneumonia',4)
In [ ]:

Result for the above model1¶

In [205]:
model1 = load_model(model_path1)
In [206]:
print_performance_results(model1)

Classes:  {'normal': 0, 'pneumonia': 1}

=================================================================

=================================================================
Accuracy:      64.56%
AUC:           62.81%
Precision:     48.09%
Recall:        31.59%
=================================================================
In [207]:
show_accuracy_charts(history1)
In [349]:
_CONFIDENCE_CRITERIA = .52
show_test_images_pred_img_panel(model1)

1/1 [==============================] - 0s 24ms/step
[[0.518805]]
1/1 [==============================] - 0s 17ms/step
[[0.51726764]]
1/1 [==============================] - 0s 31ms/step
[[0.5497466]]
1/1 [==============================] - 0s 29ms/step
[[0.32746089]]
1/1 [==============================] - 0s 20ms/step
[[0.45604503]]
1/1 [==============================] - 0s 19ms/step
[[0.37717685]]
1/1 [==============================] - 0s 17ms/step
[[0.39422864]]
1/1 [==============================] - 0s 20ms/step
[[0.3567105]]
1/1 [==============================] - 0s 21ms/step
[[0.5203447]]
1/1 [==============================] - 0s 25ms/step
[[0.41713607]]
1/1 [==============================] - 0s 34ms/step
[[0.55409855]]
1/1 [==============================] - 0s 23ms/step
[[0.5466963]]

Test Model2¶

In [290]:
#Test Model2
model2 = Sequential()
model2.add(Conv2D(filters = 32,kernel_size=(3,3),activation = 'relu' , input_shape = (IMG_HEIGHT,IMG_WIDTH,3)))
model2.add(MaxPool2D(pool_size=(2,2)))

model2.add(Conv2D(filters = 64,kernel_size=(3,3),kernel_initializer="he_uniform",activation = 'relu'))
model2.add(MaxPool2D(pool_size=(2,2)))

model2.add(Conv2D(filters = 128,kernel_size=(3,3),kernel_initializer="he_uniform",activation = 'relu'))
model2.add(MaxPool2D(pool_size=(2,2)))
model1.add(Dropout(0.25))

model2.add(Flatten())
model2.add(Dense(units=128,activation='relu'))
model2.add(Dense(units=64,activation='relu'))
model2.add(Dropout(0.50))
model2.add(Dense(units = 1 , activation = 'sigmoid'))
In [291]:
optimizer= Adam()
model2.compile(optimizer='adam', loss = 'binary_crossentropy', metrics = ['accuracy'])
model2.summary()

Model: "sequential_9"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d_28 (Conv2D)          (None, 254, 254, 32)      896       
                                                                 
 max_pooling2d_26 (MaxPoolin  (None, 127, 127, 32)     0         
 g2D)                                                            
                                                                 
 conv2d_29 (Conv2D)          (None, 125, 125, 64)      18496     
                                                                 
 max_pooling2d_27 (MaxPoolin  (None, 62, 62, 64)       0         
 g2D)                                                            
                                                                 
 conv2d_30 (Conv2D)          (None, 60, 60, 128)       73856     
                                                                 
 max_pooling2d_28 (MaxPoolin  (None, 30, 30, 128)      0         
 g2D)                                                            
                                                                 
 flatten_7 (Flatten)         (None, 115200)            0         
                                                                 
 dense_16 (Dense)            (None, 128)               14745728  
                                                                 
 dense_17 (Dense)            (None, 64)                8256      
                                                                 
 dropout_16 (Dropout)        (None, 64)                0         
                                                                 
 dense_18 (Dense)            (None, 1)                 65        
                                                                 
=================================================================
Total params: 14,847,297
Trainable params: 14,847,297
Non-trainable params: 0
_________________________________________________________________
In [292]:
model_path2 = os.path.join(saved_model_folder_path,"best_model2.h5")
print(model_path2)
early_stop2 = tf.keras.callbacks.EarlyStopping(monitor="val_accuracy",min_delta=0.01,patience=5,verbose=1,mode='auto')
checkpoint2 = ModelCheckpoint(model_path2,monitor='val_accuracy',save_best_only=True, mode='auto',verbose=1)
reduce_lr2 = ReduceLROnPlateau(monitor='accuracy',factor=0.1,patience=3,min_lr=0.00001,model='auto')
callbacks2 = [early_stop2,checkpoint2,reduce_lr2]

models_bkp/best_model2.h5
In [293]:
%%time
history2 = model2.fit_generator(generator=train_generator,epochs=50,validation_data=val_generator,validation_steps=16,workers=15,callbacks=callbacks2)

Epoch 1/50
194/195 [============================>.] - ETA: 0s - loss: 0.7188 - accuracy: 0.5318
Epoch 1: val_accuracy improved from -inf to 0.52344, saving model to models_bkp\best_model2.h5
195/195 [==============================] - 15s 69ms/step - loss: 0.7186 - accuracy: 0.5323 - val_loss: 0.6911 - val_accuracy: 0.5234 - lr: 0.0010
Epoch 2/50
194/195 [============================>.] - ETA: 0s - loss: 0.6916 - accuracy: 0.5344
Epoch 2: val_accuracy improved from 0.52344 to 0.54297, saving model to models_bkp\best_model2.h5
195/195 [==============================] - 14s 67ms/step - loss: 0.6916 - accuracy: 0.5342 - val_loss: 0.6884 - val_accuracy: 0.5430 - lr: 0.0010
Epoch 3/50
194/195 [============================>.] - ETA: 0s - loss: 0.6907 - accuracy: 0.5389
Epoch 3: val_accuracy did not improve from 0.54297
195/195 [==============================] - 14s 66ms/step - loss: 0.6908 - accuracy: 0.5387 - val_loss: 0.6931 - val_accuracy: 0.5156 - lr: 0.0010
Epoch 4/50
194/195 [============================>.] - ETA: 0s - loss: 0.6923 - accuracy: 0.5408
Epoch 4: val_accuracy did not improve from 0.54297
195/195 [==============================] - 14s 68ms/step - loss: 0.6923 - accuracy: 0.5410 - val_loss: 0.6910 - val_accuracy: 0.5234 - lr: 0.0010
Epoch 5/50
195/195 [==============================] - ETA: 0s - loss: 0.6891 - accuracy: 0.5432
Epoch 5: val_accuracy improved from 0.54297 to 0.54688, saving model to models_bkp\best_model2.h5
195/195 [==============================] - 15s 72ms/step - loss: 0.6891 - accuracy: 0.5432 - val_loss: 0.6909 - val_accuracy: 0.5469 - lr: 0.0010
Epoch 6/50
195/195 [==============================] - ETA: 0s - loss: 0.6902 - accuracy: 0.5371
Epoch 6: val_accuracy improved from 0.54688 to 0.58984, saving model to models_bkp\best_model2.h5
195/195 [==============================] - 16s 76ms/step - loss: 0.6902 - accuracy: 0.5371 - val_loss: 0.6833 - val_accuracy: 0.5898 - lr: 0.0010
Epoch 7/50
195/195 [==============================] - ETA: 0s - loss: 0.6901 - accuracy: 0.5320
Epoch 7: val_accuracy did not improve from 0.58984
195/195 [==============================] - 15s 73ms/step - loss: 0.6901 - accuracy: 0.5320 - val_loss: 0.6838 - val_accuracy: 0.5586 - lr: 0.0010
Epoch 8/50
195/195 [==============================] - ETA: 0s - loss: 0.6903 - accuracy: 0.5413
Epoch 8: val_accuracy did not improve from 0.58984
195/195 [==============================] - 17s 80ms/step - loss: 0.6903 - accuracy: 0.5413 - val_loss: 0.6932 - val_accuracy: 0.5234 - lr: 0.0010
Epoch 9/50
195/195 [==============================] - ETA: 0s - loss: 0.6876 - accuracy: 0.5416
Epoch 9: val_accuracy did not improve from 0.58984
195/195 [==============================] - 15s 75ms/step - loss: 0.6876 - accuracy: 0.5416 - val_loss: 0.6884 - val_accuracy: 0.5391 - lr: 1.0000e-04
Epoch 10/50
195/195 [==============================] - ETA: 0s - loss: 0.6863 - accuracy: 0.5512
Epoch 10: val_accuracy did not improve from 0.58984
195/195 [==============================] - 16s 76ms/step - loss: 0.6863 - accuracy: 0.5512 - val_loss: 0.6895 - val_accuracy: 0.5469 - lr: 1.0000e-04
Epoch 11/50
194/195 [============================>.] - ETA: 0s - loss: 0.6870 - accuracy: 0.5476
Epoch 11: val_accuracy did not improve from 0.58984
195/195 [==============================] - 16s 75ms/step - loss: 0.6870 - accuracy: 0.5480 - val_loss: 0.6816 - val_accuracy: 0.5547 - lr: 1.0000e-04
Epoch 11: early stopping
CPU times: total: 17min 7s
Wall time: 2min 45s
In [272]:
model2 = load_model(model_path2)

Result for the above model2¶

In [111]:
print_performance_results(model2)

Classes:  {'normal': 0, 'pneumonia': 1}

=================================================================

=================================================================
Accuracy:      62.61%
AUC:           66.44%
Precision:     47.20%
Recall:        45.91%
=================================================================
In [273]:
show_accuracy_charts(history2)
In [350]:
_CONFIDENCE_CRITERIA = .52
show_test_images_pred_img_panel(model2)

1/1 [==============================] - 0s 303ms/step
[[0.43902734]]
1/1 [==============================] - 0s 16ms/step
[[0.4632777]]
1/1 [==============================] - 0s 16ms/step
[[0.45219988]]
1/1 [==============================] - 0s 31ms/step
[[0.43675944]]
1/1 [==============================] - 0s 16ms/step
[[0.44630206]]
1/1 [==============================] - 0s 15ms/step
[[0.43865]]
1/1 [==============================] - 0s 16ms/step
[[0.45661345]]
1/1 [==============================] - 0s 16ms/step
[[0.44729123]]
1/1 [==============================] - 0s 19ms/step
[[0.4572494]]
1/1 [==============================] - 0s 19ms/step
[[0.4433138]]
1/1 [==============================] - 0s 16ms/step
[[0.4882266]]
1/1 [==============================] - 0s 19ms/step
[[0.46829396]]

Using MobileNetV2¶

MobileNet was trained to classify images with 1000 class of images and at the end it have dense layer with 1000 nodes or activation layers and we want to take it off¶

"include_top" is kept to False as we do not want to include dense layers with 1000 nodes and we want to train our model on the top of this pretrained model and should not include 1000 classes to classify in our model "pooling" we have set as "avg" so that allow us to have one vector output

In [351]:
mobilenet = tf.keras.applications.MobileNetV2(
    input_shape=(IMG_HEIGHT,IMG_WIDTH,3),
    include_top=False,
    weights="imagenet",
    pooling='avg')
mobilenet.trainable = False #Keeps the weights being trained

WARNING:tensorflow:`input_shape` is undefined or non-square, or `rows` is not in [96, 128, 160, 192, 224]. Weights for input shape (224, 224) will be loaded as the default.
In [352]:
inputs = tf.keras.Input(shape=(IMG_HEIGHT,IMG_WIDTH,3))
pretrained_model = mobilenet(inputs,training=False)
dense = tf.keras.layers.Dense(1024,activation='relu')(pretrained_model) 
outputs = tf.keras.layers.Dense(1,activation="sigmoid")(dense)
model3 = tf.keras.Model(inputs,outputs)
print(model3.summary())

Model: "model"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_2 (InputLayer)        [(None, 256, 256, 3)]     0         
                                                                 
 mobilenetv2_1.00_224 (Funct  (None, 1280)             2257984   
 ional)                                                          
                                                                 
 dense_19 (Dense)            (None, 1024)              1311744   
                                                                 
 dense_20 (Dense)            (None, 1)                 1025      
                                                                 
=================================================================
Total params: 3,570,753
Trainable params: 1,312,769
Non-trainable params: 2,257,984
_________________________________________________________________
None
In [353]:
model_path3 = os.path.join(saved_model_folder_path,"best_model3.h5")
print(model_path3)
early_stop3 = tf.keras.callbacks.EarlyStopping(monitor="val_accuracy",min_delta=0.01,patience=5,verbose=1,mode='auto')
checkpoint3 = ModelCheckpoint(model_path3,monitor='val_accuracy',save_best_only=True, mode='auto',verbose=1)
reduce_lr3 = ReduceLROnPlateau(monitor='accuracy',factor=0.1,patience=3,min_lr=0.00001,model='auto')
callbacks3 = [early_stop3,checkpoint3,reduce_lr3]

models_bkp/best_model3.h5
In [355]:
EPOCHS=50
BATCH_SIZE=35
model3.compile(
    optimizer='adam',
    loss='binary_crossentropy',
    metrics=['accuracy',tf.keras.metrics.AUC(name='auc')]
)
history3 = model3.fit(
    train_generator,
    validation_data=val_generator,
    batch_size=BATCH_SIZE,
    epochs=EPOCHS,
    callbacks=callbacks3
)

Epoch 1/50
195/195 [==============================] - ETA: 0s - loss: 0.6335 - accuracy: 0.6563 - auc: 0.7034
Epoch 1: val_accuracy improved from -inf to 0.66089, saving model to models_bkp\best_model3.h5
195/195 [==============================] - 63s 306ms/step - loss: 0.6335 - accuracy: 0.6563 - auc: 0.7034 - val_loss: 0.6011 - val_accuracy: 0.6609 - val_auc: 0.7611 - lr: 0.0010
Epoch 2/50
195/195 [==============================] - ETA: 0s - loss: 0.6106 - accuracy: 0.6627 - auc: 0.7266
Epoch 2: val_accuracy improved from 0.66089 to 0.68882, saving model to models_bkp\best_model3.h5
195/195 [==============================] - 58s 295ms/step - loss: 0.6106 - accuracy: 0.6627 - auc: 0.7266 - val_loss: 0.5834 - val_accuracy: 0.6888 - val_auc: 0.7711 - lr: 0.0010
Epoch 3/50
195/195 [==============================] - ETA: 0s - loss: 0.6038 - accuracy: 0.6762 - auc: 0.7355
Epoch 3: val_accuracy did not improve from 0.68882
195/195 [==============================] - 57s 292ms/step - loss: 0.6038 - accuracy: 0.6762 - auc: 0.7355 - val_loss: 0.5852 - val_accuracy: 0.6840 - val_auc: 0.7752 - lr: 0.0010
Epoch 4/50
195/195 [==============================] - ETA: 0s - loss: 0.5781 - accuracy: 0.6984 - auc: 0.7627
Epoch 4: val_accuracy did not improve from 0.68882
195/195 [==============================] - 58s 298ms/step - loss: 0.5781 - accuracy: 0.6984 - auc: 0.7627 - val_loss: 0.6396 - val_accuracy: 0.6349 - val_auc: 0.7804 - lr: 0.0010
Epoch 5/50
195/195 [==============================] - ETA: 0s - loss: 0.5868 - accuracy: 0.6858 - auc: 0.7526
Epoch 5: val_accuracy improved from 0.68882 to 0.71291, saving model to models_bkp\best_model3.h5
195/195 [==============================] - 58s 296ms/step - loss: 0.5868 - accuracy: 0.6858 - auc: 0.7526 - val_loss: 0.5719 - val_accuracy: 0.7129 - val_auc: 0.7800 - lr: 0.0010
Epoch 6/50
195/195 [==============================] - ETA: 0s - loss: 0.5772 - accuracy: 0.6964 - auc: 0.7627
Epoch 6: val_accuracy did not improve from 0.71291
195/195 [==============================] - 57s 291ms/step - loss: 0.5772 - accuracy: 0.6964 - auc: 0.7627 - val_loss: 0.5793 - val_accuracy: 0.6956 - val_auc: 0.7792 - lr: 0.0010
Epoch 7/50
195/195 [==============================] - ETA: 0s - loss: 0.5761 - accuracy: 0.6951 - auc: 0.7634
Epoch 7: val_accuracy did not improve from 0.71291
195/195 [==============================] - 55s 283ms/step - loss: 0.5761 - accuracy: 0.6951 - auc: 0.7634 - val_loss: 0.6133 - val_accuracy: 0.6657 - val_auc: 0.7864 - lr: 0.0010
Epoch 8/50
195/195 [==============================] - ETA: 0s - loss: 0.5515 - accuracy: 0.7244 - auc: 0.7895
Epoch 8: val_accuracy did not improve from 0.71291
195/195 [==============================] - 50s 257ms/step - loss: 0.5515 - accuracy: 0.7244 - auc: 0.7895 - val_loss: 0.5858 - val_accuracy: 0.6821 - val_auc: 0.7857 - lr: 1.0000e-04
Epoch 9/50
195/195 [==============================] - ETA: 0s - loss: 0.5530 - accuracy: 0.7151 - auc: 0.7873
Epoch 9: val_accuracy did not improve from 0.71291
195/195 [==============================] - 50s 259ms/step - loss: 0.5530 - accuracy: 0.7151 - auc: 0.7873 - val_loss: 0.5928 - val_accuracy: 0.6802 - val_auc: 0.7862 - lr: 1.0000e-04
Epoch 10/50
195/195 [==============================] - ETA: 0s - loss: 0.5551 - accuracy: 0.7205 - auc: 0.7852
Epoch 10: val_accuracy did not improve from 0.71291
195/195 [==============================] - 48s 247ms/step - loss: 0.5551 - accuracy: 0.7205 - auc: 0.7852 - val_loss: 0.5759 - val_accuracy: 0.6908 - val_auc: 0.7896 - lr: 1.0000e-04
Epoch 10: early stopping

Results¶

In [356]:
model3 = load_model(model_path3)
In [357]:
print_performance_results(model3)

Classes:  {'normal': 0, 'pneumonia': 1}

=================================================================

=================================================================
Accuracy:      58.01%
AUC:           69.46%
Precision:     46.74%
Recall:        43.51%
=================================================================
In [358]:
show_accuracy_charts(history3)
In [359]:
_CONFIDENCE_CRITERIA = .52
show_test_images_pred_img_panel(model3)

1/1 [==============================] - 1s 583ms/step
[[0.20442389]]
1/1 [==============================] - 0s 31ms/step
[[0.33535773]]
1/1 [==============================] - 0s 31ms/step
[[0.6734893]]
1/1 [==============================] - 0s 22ms/step
[[0.18333726]]
1/1 [==============================] - 0s 20ms/step
[[0.48843879]]
1/1 [==============================] - 0s 16ms/step
[[0.10179777]]
1/1 [==============================] - 0s 34ms/step
[[0.311572]]
1/1 [==============================] - 0s 18ms/step
[[0.31551057]]
1/1 [==============================] - 0s 15ms/step
[[0.5656961]]
1/1 [==============================] - 0s 20ms/step
[[0.4989342]]
1/1 [==============================] - 0s 16ms/step
[[0.3715186]]
1/1 [==============================] - 0s 19ms/step
[[0.537735]]
In [ ]:

Using VGG 16¶

In [360]:
#importing the VGG16 library as shown below and add preprocessing layer to the front of the VGG
#We will use imagenet weights. include_top ensure we dont get first and the last layer 
vgg = VGG16(input_shape=[IMG_HEIGHT,IMG_WIDTH]+[3],weights='imagenet',include_top=False)
In [361]:
#ensuring all the layer dont get trained
for layer in vgg.layers:
    layer.trainable = False
In [362]:
vgg16_x = Flatten()(vgg.output)
vgg16_pred = Dense(1,activation="sigmoid")(vgg16_x)
model4 = tf.keras.Model(inputs=vgg.input,outputs=vgg16_pred)
In [363]:
model4.summary()

Model: "model_1"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_3 (InputLayer)        [(None, 256, 256, 3)]     0         
                                                                 
 block1_conv1 (Conv2D)       (None, 256, 256, 64)      1792      
                                                                 
 block1_conv2 (Conv2D)       (None, 256, 256, 64)      36928     
                                                                 
 block1_pool (MaxPooling2D)  (None, 128, 128, 64)      0         
                                                                 
 block2_conv1 (Conv2D)       (None, 128, 128, 128)     73856     
                                                                 
 block2_conv2 (Conv2D)       (None, 128, 128, 128)     147584    
                                                                 
 block2_pool (MaxPooling2D)  (None, 64, 64, 128)       0         
                                                                 
 block3_conv1 (Conv2D)       (None, 64, 64, 256)       295168    
                                                                 
 block3_conv2 (Conv2D)       (None, 64, 64, 256)       590080    
                                                                 
 block3_conv3 (Conv2D)       (None, 64, 64, 256)       590080    
                                                                 
 block3_pool (MaxPooling2D)  (None, 32, 32, 256)       0         
                                                                 
 block4_conv1 (Conv2D)       (None, 32, 32, 512)       1180160   
                                                                 
 block4_conv2 (Conv2D)       (None, 32, 32, 512)       2359808   
                                                                 
 block4_conv3 (Conv2D)       (None, 32, 32, 512)       2359808   
                                                                 
 block4_pool (MaxPooling2D)  (None, 16, 16, 512)       0         
                                                                 
 block5_conv1 (Conv2D)       (None, 16, 16, 512)       2359808   
                                                                 
 block5_conv2 (Conv2D)       (None, 16, 16, 512)       2359808   
                                                                 
 block5_conv3 (Conv2D)       (None, 16, 16, 512)       2359808   
                                                                 
 block5_pool (MaxPooling2D)  (None, 8, 8, 512)         0         
                                                                 
 flatten_8 (Flatten)         (None, 32768)             0         
                                                                 
 dense_21 (Dense)            (None, 1)                 32769     
                                                                 
=================================================================
Total params: 14,747,457
Trainable params: 32,769
Non-trainable params: 14,714,688
_________________________________________________________________
In [364]:
model_path4 = os.path.join(saved_model_folder_path,"best_model4.h5")
print(model_path4)
early_stop4 = tf.keras.callbacks.EarlyStopping(monitor="val_accuracy",min_delta=0.01,patience=10,verbose=1,mode='auto')
checkpoint4 = ModelCheckpoint(model_path4,monitor='val_accuracy',save_best_only=True, mode='auto',verbose=1)
reduce_lr4 = ReduceLROnPlateau(monitor='accuracy',factor=0.1,patience=3,min_lr=0.00001,model='auto')
callbacks4 = [early_stop4,checkpoint4,reduce_lr4]

models_bkp/best_model4.h5
In [366]:
model4.compile(
    optimizer='adam',
    loss='binary_crossentropy',
    metrics=['accuracy',tf.keras.metrics.AUC(name='auc')]
)
history4 = model4.fit_generator(
    train_generator,
    validation_data= val_generator,
    epochs=EPOCHS,
    callbacks=callbacks4
)

Epoch 1/50
195/195 [==============================] - ETA: 0s - loss: 0.7089 - accuracy: 0.6354 - auc: 0.6763
Epoch 1: val_accuracy improved from -inf to 0.67726, saving model to models_bkp\best_model4.h5
195/195 [==============================] - 55s 280ms/step - loss: 0.7089 - accuracy: 0.6354 - auc: 0.6763 - val_loss: 0.6121 - val_accuracy: 0.6773 - val_auc: 0.7521 - lr: 0.0010
Epoch 2/50
195/195 [==============================] - ETA: 0s - loss: 0.6689 - accuracy: 0.6624 - auc: 0.7205
Epoch 2: val_accuracy did not improve from 0.67726
195/195 [==============================] - 57s 292ms/step - loss: 0.6689 - accuracy: 0.6624 - auc: 0.7205 - val_loss: 0.6603 - val_accuracy: 0.6657 - val_auc: 0.7763 - lr: 0.0010
Epoch 3/50
195/195 [==============================] - ETA: 0s - loss: 0.6771 - accuracy: 0.6627 - auc: 0.7186
Epoch 3: val_accuracy improved from 0.67726 to 0.68304, saving model to models_bkp\best_model4.h5
195/195 [==============================] - 54s 278ms/step - loss: 0.6771 - accuracy: 0.6627 - auc: 0.7186 - val_loss: 0.6154 - val_accuracy: 0.6830 - val_auc: 0.7766 - lr: 0.0010
Epoch 4/50
195/195 [==============================] - ETA: 0s - loss: 0.6342 - accuracy: 0.6894 - auc: 0.7523
Epoch 4: val_accuracy improved from 0.68304 to 0.70520, saving model to models_bkp\best_model4.h5
195/195 [==============================] - 52s 264ms/step - loss: 0.6342 - accuracy: 0.6894 - auc: 0.7523 - val_loss: 0.5813 - val_accuracy: 0.7052 - val_auc: 0.7758 - lr: 0.0010
Epoch 5/50
195/195 [==============================] - ETA: 0s - loss: 0.6170 - accuracy: 0.6878 - auc: 0.7583
Epoch 5: val_accuracy did not improve from 0.70520
195/195 [==============================] - 52s 265ms/step - loss: 0.6170 - accuracy: 0.6878 - auc: 0.7583 - val_loss: 0.5847 - val_accuracy: 0.7023 - val_auc: 0.7708 - lr: 0.0010
Epoch 6/50
195/195 [==============================] - ETA: 0s - loss: 0.6337 - accuracy: 0.6942 - auc: 0.7522
Epoch 6: val_accuracy did not improve from 0.70520
195/195 [==============================] - 52s 266ms/step - loss: 0.6337 - accuracy: 0.6942 - auc: 0.7522 - val_loss: 0.7294 - val_accuracy: 0.6358 - val_auc: 0.7689 - lr: 0.0010
Epoch 7/50
195/195 [==============================] - ETA: 0s - loss: 0.6147 - accuracy: 0.6984 - auc: 0.7623
Epoch 7: val_accuracy did not improve from 0.70520
195/195 [==============================] - 51s 262ms/step - loss: 0.6147 - accuracy: 0.6984 - auc: 0.7623 - val_loss: 0.7512 - val_accuracy: 0.6108 - val_auc: 0.7718 - lr: 0.0010
Epoch 8/50
195/195 [==============================] - ETA: 0s - loss: 0.6751 - accuracy: 0.6736 - auc: 0.7400
Epoch 8: val_accuracy did not improve from 0.70520
195/195 [==============================] - 51s 262ms/step - loss: 0.6751 - accuracy: 0.6736 - auc: 0.7400 - val_loss: 0.7566 - val_accuracy: 0.6262 - val_auc: 0.7727 - lr: 0.0010
Epoch 9/50
195/195 [==============================] - ETA: 0s - loss: 0.6201 - accuracy: 0.6951 - auc: 0.7613
Epoch 9: val_accuracy did not improve from 0.70520
195/195 [==============================] - 52s 265ms/step - loss: 0.6201 - accuracy: 0.6951 - auc: 0.7613 - val_loss: 0.6643 - val_accuracy: 0.6898 - val_auc: 0.7791 - lr: 0.0010
Epoch 10/50
195/195 [==============================] - ETA: 0s - loss: 0.6239 - accuracy: 0.7025 - auc: 0.7675
Epoch 10: val_accuracy did not improve from 0.70520
195/195 [==============================] - 51s 261ms/step - loss: 0.6239 - accuracy: 0.7025 - auc: 0.7675 - val_loss: 0.6270 - val_accuracy: 0.6811 - val_auc: 0.7761 - lr: 0.0010
Epoch 11/50
195/195 [==============================] - ETA: 0s - loss: 0.6314 - accuracy: 0.7025 - auc: 0.7654
Epoch 11: val_accuracy improved from 0.70520 to 0.71098, saving model to models_bkp\best_model4.h5
195/195 [==============================] - 57s 293ms/step - loss: 0.6314 - accuracy: 0.7025 - auc: 0.7654 - val_loss: 0.6304 - val_accuracy: 0.7110 - val_auc: 0.7773 - lr: 0.0010
Epoch 12/50
195/195 [==============================] - ETA: 0s - loss: 0.6328 - accuracy: 0.6878 - auc: 0.7645
Epoch 12: val_accuracy improved from 0.71098 to 0.73025, saving model to models_bkp\best_model4.h5
195/195 [==============================] - 58s 297ms/step - loss: 0.6328 - accuracy: 0.6878 - auc: 0.7645 - val_loss: 0.5803 - val_accuracy: 0.7303 - val_auc: 0.7880 - lr: 0.0010
Epoch 13/50
195/195 [==============================] - ETA: 0s - loss: 0.5933 - accuracy: 0.7086 - auc: 0.7832
Epoch 13: val_accuracy did not improve from 0.73025
195/195 [==============================] - 51s 259ms/step - loss: 0.5933 - accuracy: 0.7086 - auc: 0.7832 - val_loss: 0.6143 - val_accuracy: 0.7091 - val_auc: 0.7818 - lr: 0.0010
Epoch 14/50
195/195 [==============================] - ETA: 0s - loss: 0.6103 - accuracy: 0.7112 - auc: 0.7806
Epoch 14: val_accuracy did not improve from 0.73025
195/195 [==============================] - 50s 256ms/step - loss: 0.6103 - accuracy: 0.7112 - auc: 0.7806 - val_loss: 0.7682 - val_accuracy: 0.6551 - val_auc: 0.7859 - lr: 0.0010
Epoch 15/50
195/195 [==============================] - ETA: 0s - loss: 0.5984 - accuracy: 0.7151 - auc: 0.7827
Epoch 15: val_accuracy did not improve from 0.73025
195/195 [==============================] - 51s 262ms/step - loss: 0.5984 - accuracy: 0.7151 - auc: 0.7827 - val_loss: 0.7217 - val_accuracy: 0.6368 - val_auc: 0.7853 - lr: 0.0010
Epoch 16/50
195/195 [==============================] - ETA: 0s - loss: 0.6101 - accuracy: 0.7119 - auc: 0.7822
Epoch 16: val_accuracy did not improve from 0.73025
195/195 [==============================] - 53s 270ms/step - loss: 0.6101 - accuracy: 0.7119 - auc: 0.7822 - val_loss: 0.7518 - val_accuracy: 0.6532 - val_auc: 0.7598 - lr: 0.0010
Epoch 17/50
195/195 [==============================] - ETA: 0s - loss: 0.6083 - accuracy: 0.7154 - auc: 0.7834
Epoch 17: val_accuracy did not improve from 0.73025
195/195 [==============================] - 52s 264ms/step - loss: 0.6083 - accuracy: 0.7154 - auc: 0.7834 - val_loss: 0.6828 - val_accuracy: 0.6618 - val_auc: 0.7723 - lr: 0.0010
Epoch 18/50
195/195 [==============================] - ETA: 0s - loss: 0.6283 - accuracy: 0.7019 - auc: 0.7729
Epoch 18: val_accuracy did not improve from 0.73025
195/195 [==============================] - 51s 262ms/step - loss: 0.6283 - accuracy: 0.7019 - auc: 0.7729 - val_loss: 0.7074 - val_accuracy: 0.6561 - val_auc: 0.7733 - lr: 0.0010
Epoch 19/50
195/195 [==============================] - ETA: 0s - loss: 0.5923 - accuracy: 0.7225 - auc: 0.7882
Epoch 19: val_accuracy did not improve from 0.73025
195/195 [==============================] - 51s 264ms/step - loss: 0.5923 - accuracy: 0.7225 - auc: 0.7882 - val_loss: 0.5981 - val_accuracy: 0.7081 - val_auc: 0.7762 - lr: 0.0010
Epoch 20/50
195/195 [==============================] - ETA: 0s - loss: 0.5752 - accuracy: 0.7221 - auc: 0.7974
Epoch 20: val_accuracy did not improve from 0.73025
195/195 [==============================] - 53s 269ms/step - loss: 0.5752 - accuracy: 0.7221 - auc: 0.7974 - val_loss: 0.7949 - val_accuracy: 0.6272 - val_auc: 0.7835 - lr: 0.0010
Epoch 21/50
195/195 [==============================] - ETA: 0s - loss: 0.5853 - accuracy: 0.7176 - auc: 0.7946
Epoch 21: val_accuracy did not improve from 0.73025
195/195 [==============================] - 50s 256ms/step - loss: 0.5853 - accuracy: 0.7176 - auc: 0.7946 - val_loss: 0.5940 - val_accuracy: 0.7139 - val_auc: 0.7857 - lr: 0.0010
Epoch 22/50
195/195 [==============================] - ETA: 0s - loss: 0.6360 - accuracy: 0.7128 - auc: 0.7759
Epoch 22: val_accuracy did not improve from 0.73025
195/195 [==============================] - 49s 252ms/step - loss: 0.6360 - accuracy: 0.7128 - auc: 0.7759 - val_loss: 0.8497 - val_accuracy: 0.6252 - val_auc: 0.7675 - lr: 0.0010
Epoch 22: early stopping
In [367]:
model4 = load_model(model_path4)

Model4 Result¶

In [368]:
print_performance_results(model4)

Classes:  {'normal': 0, 'pneumonia': 1}

=================================================================

=================================================================
Accuracy:      61.52%
AUC:           71.00%
Precision:     46.99%
Recall:        40.79%
=================================================================
In [369]:
show_accuracy_charts(history4)
In [370]:
_CONFIDENCE_CRITERIA = .52
show_test_images_pred_img_panel(model4)

1/1 [==============================] - 1s 503ms/step
[[0.15321851]]
1/1 [==============================] - 0s 16ms/step
[[0.60776967]]
1/1 [==============================] - 0s 30ms/step
[[0.78351]]
1/1 [==============================] - 0s 32ms/step
[[0.05015829]]
1/1 [==============================] - 0s 31ms/step
[[0.36371034]]
1/1 [==============================] - 0s 32ms/step
[[0.06952075]]
1/1 [==============================] - 0s 32ms/step
[[0.448549]]
1/1 [==============================] - 0s 35ms/step
[[0.23729156]]
1/1 [==============================] - 0s 16ms/step
[[0.709824]]
1/1 [==============================] - 0s 16ms/step
[[0.38517132]]
1/1 [==============================] - 0s 19ms/step
[[0.38598502]]
1/1 [==============================] - 0s 32ms/step
[[0.7682301]]

Model 5¶

Using VGG 16 _2nd config¶

In [538]:
IMG_WIDTH_VGG16 = 350
IMG_HEIGHT_VGG16 = 350
#importing the VGG16 library as shown below and add preprocessing layer to the front of the VGG
#We will use imagenet weights. include_top ensure we dont get first and the last layer 
vgg2 = VGG16(input_shape=[IMG_HEIGHT_VGG16,IMG_WIDTH_VGG16]+[3],weights='imagenet',include_top=False)
#ensuring all the layer dont get trained
for layer in vgg2.layers:
    layer.trainable = False
In [540]:
vgg16_x2 = Flatten()(vgg2.output)
vgg16_pred2 = Dense(1,activation="sigmoid")(vgg16_x2)
model5 = tf.keras.Model(inputs=vgg2.input,outputs=vgg16_pred2)
In [542]:
model5.summary()

Model: "model_4"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_8 (InputLayer)        [(None, 350, 350, 3)]     0         
                                                                 
 block1_conv1 (Conv2D)       (None, 350, 350, 64)      1792      
                                                                 
 block1_conv2 (Conv2D)       (None, 350, 350, 64)      36928     
                                                                 
 block1_pool (MaxPooling2D)  (None, 175, 175, 64)      0         
                                                                 
 block2_conv1 (Conv2D)       (None, 175, 175, 128)     73856     
                                                                 
 block2_conv2 (Conv2D)       (None, 175, 175, 128)     147584    
                                                                 
 block2_pool (MaxPooling2D)  (None, 87, 87, 128)       0         
                                                                 
 block3_conv1 (Conv2D)       (None, 87, 87, 256)       295168    
                                                                 
 block3_conv2 (Conv2D)       (None, 87, 87, 256)       590080    
                                                                 
 block3_conv3 (Conv2D)       (None, 87, 87, 256)       590080    
                                                                 
 block3_pool (MaxPooling2D)  (None, 43, 43, 256)       0         
                                                                 
 block4_conv1 (Conv2D)       (None, 43, 43, 512)       1180160   
                                                                 
 block4_conv2 (Conv2D)       (None, 43, 43, 512)       2359808   
                                                                 
 block4_conv3 (Conv2D)       (None, 43, 43, 512)       2359808   
                                                                 
 block4_pool (MaxPooling2D)  (None, 21, 21, 512)       0         
                                                                 
 block5_conv1 (Conv2D)       (None, 21, 21, 512)       2359808   
                                                                 
 block5_conv2 (Conv2D)       (None, 21, 21, 512)       2359808   
                                                                 
 block5_conv3 (Conv2D)       (None, 21, 21, 512)       2359808   
                                                                 
 block5_pool (MaxPooling2D)  (None, 10, 10, 512)       0         
                                                                 
 flatten_11 (Flatten)        (None, 51200)             0         
                                                                 
 dense_24 (Dense)            (None, 1)                 51201     
                                                                 
=================================================================
Total params: 14,765,889
Trainable params: 51,201
Non-trainable params: 14,714,688
_________________________________________________________________
In [543]:
model_path5 = os.path.join(saved_model_folder_path,"best_model5.h5")
print(model_path5)
early_stop5 = tf.keras.callbacks.EarlyStopping(monitor="val_accuracy",min_delta=0.01,patience=10,verbose=1,mode='auto')
checkpoint5 = ModelCheckpoint(model_path5,monitor='val_accuracy',save_best_only=True, mode='auto',verbose=1)
reduce_lr5 = ReduceLROnPlateau(monitor='accuracy',factor=0.1,patience=3,min_lr=0.00001,model='auto')
callbacks5 = [early_stop5,checkpoint5,reduce_lr5]

models_bkp/best_model5.h5
In [546]:
model5.compile(
    optimizer='adam',
    loss='binary_crossentropy',
    metrics=['accuracy',tf.keras.metrics.AUC(name='auc')]
)
history5 = model5.fit_generator(
    train_generator_vgg16,
    validation_data= val_generator_vgg16,
    epochs=EPOCHS,
    callbacks=callbacks5
)

Epoch 1/50
195/195 [==============================] - ETA: 0s - loss: 0.8236 - accuracy: 0.6065 - auc: 0.6394
Epoch 1: val_accuracy improved from -inf to 0.67919, saving model to models_bkp\best_model5.h5
195/195 [==============================] - 118s 593ms/step - loss: 0.8236 - accuracy: 0.6065 - auc: 0.6394 - val_loss: 0.5968 - val_accuracy: 0.6792 - val_auc: 0.7561 - lr: 0.0010
Epoch 2/50
195/195 [==============================] - ETA: 0s - loss: 0.6807 - accuracy: 0.6640 - auc: 0.7237
Epoch 2: val_accuracy improved from 0.67919 to 0.68882, saving model to models_bkp\best_model5.h5
195/195 [==============================] - 101s 518ms/step - loss: 0.6807 - accuracy: 0.6640 - auc: 0.7237 - val_loss: 0.5995 - val_accuracy: 0.6888 - val_auc: 0.7617 - lr: 0.0010
Epoch 3/50
195/195 [==============================] - ETA: 0s - loss: 0.6307 - accuracy: 0.6907 - auc: 0.7575
Epoch 3: val_accuracy did not improve from 0.68882
195/195 [==============================] - 102s 523ms/step - loss: 0.6307 - accuracy: 0.6907 - auc: 0.7575 - val_loss: 0.6843 - val_accuracy: 0.6599 - val_auc: 0.7667 - lr: 0.0010
Epoch 4/50
195/195 [==============================] - ETA: 0s - loss: 0.6563 - accuracy: 0.6926 - auc: 0.7526
Epoch 4: val_accuracy improved from 0.68882 to 0.69750, saving model to models_bkp\best_model5.h5
195/195 [==============================] - 102s 523ms/step - loss: 0.6563 - accuracy: 0.6926 - auc: 0.7526 - val_loss: 0.6186 - val_accuracy: 0.6975 - val_auc: 0.7634 - lr: 0.0010
Epoch 5/50
195/195 [==============================] - ETA: 0s - loss: 0.5630 - accuracy: 0.7273 - auc: 0.7978
Epoch 5: val_accuracy did not improve from 0.69750
195/195 [==============================] - 104s 531ms/step - loss: 0.5630 - accuracy: 0.7273 - auc: 0.7978 - val_loss: 0.7097 - val_accuracy: 0.6484 - val_auc: 0.7816 - lr: 0.0010
Epoch 6/50
195/195 [==============================] - ETA: 0s - loss: 0.5788 - accuracy: 0.7199 - auc: 0.7945
Epoch 6: val_accuracy did not improve from 0.69750
195/195 [==============================] - 101s 517ms/step - loss: 0.5788 - accuracy: 0.7199 - auc: 0.7945 - val_loss: 0.6375 - val_accuracy: 0.6965 - val_auc: 0.7697 - lr: 0.0010
Epoch 7/50
195/195 [==============================] - ETA: 0s - loss: 0.5698 - accuracy: 0.7234 - auc: 0.8024
Epoch 7: val_accuracy did not improve from 0.69750
195/195 [==============================] - 99s 509ms/step - loss: 0.5698 - accuracy: 0.7234 - auc: 0.8024 - val_loss: 0.6709 - val_accuracy: 0.6696 - val_auc: 0.7856 - lr: 0.0010
Epoch 8/50
195/195 [==============================] - ETA: 0s - loss: 0.5948 - accuracy: 0.7112 - auc: 0.7906
Epoch 8: val_accuracy did not improve from 0.69750
195/195 [==============================] - 100s 512ms/step - loss: 0.5948 - accuracy: 0.7112 - auc: 0.7906 - val_loss: 0.7687 - val_accuracy: 0.6407 - val_auc: 0.7693 - lr: 0.0010
Epoch 9/50
195/195 [==============================] - ETA: 0s - loss: 0.4635 - accuracy: 0.7838 - auc: 0.8622
Epoch 9: val_accuracy improved from 0.69750 to 0.70520, saving model to models_bkp\best_model5.h5
195/195 [==============================] - 94s 483ms/step - loss: 0.4635 - accuracy: 0.7838 - auc: 0.8622 - val_loss: 0.6131 - val_accuracy: 0.7052 - val_auc: 0.7687 - lr: 1.0000e-04
Epoch 10/50
195/195 [==============================] - ETA: 0s - loss: 0.4540 - accuracy: 0.7809 - auc: 0.8663
Epoch 10: val_accuracy did not improve from 0.70520
195/195 [==============================] - 98s 501ms/step - loss: 0.4540 - accuracy: 0.7809 - auc: 0.8663 - val_loss: 0.6146 - val_accuracy: 0.6994 - val_auc: 0.7725 - lr: 1.0000e-04
Epoch 11/50
195/195 [==============================] - ETA: 0s - loss: 0.4610 - accuracy: 0.7751 - auc: 0.8601
Epoch 11: val_accuracy improved from 0.70520 to 0.71484, saving model to models_bkp\best_model5.h5
195/195 [==============================] - 102s 521ms/step - loss: 0.4610 - accuracy: 0.7751 - auc: 0.8601 - val_loss: 0.5924 - val_accuracy: 0.7148 - val_auc: 0.7746 - lr: 1.0000e-04
Epoch 12/50
195/195 [==============================] - ETA: 0s - loss: 0.4563 - accuracy: 0.7864 - auc: 0.8659
Epoch 12: val_accuracy did not improve from 0.71484
195/195 [==============================] - 101s 518ms/step - loss: 0.4563 - accuracy: 0.7864 - auc: 0.8659 - val_loss: 0.6006 - val_accuracy: 0.7013 - val_auc: 0.7745 - lr: 1.0000e-04
Epoch 13/50
195/195 [==============================] - ETA: 0s - loss: 0.4463 - accuracy: 0.7928 - auc: 0.8727
Epoch 13: val_accuracy did not improve from 0.71484
195/195 [==============================] - 104s 532ms/step - loss: 0.4463 - accuracy: 0.7928 - auc: 0.8727 - val_loss: 0.5908 - val_accuracy: 0.7091 - val_auc: 0.7741 - lr: 1.0000e-04
Epoch 14/50
195/195 [==============================] - ETA: 0s - loss: 0.4383 - accuracy: 0.7931 - auc: 0.8786
Epoch 14: val_accuracy did not improve from 0.71484
195/195 [==============================] - 94s 483ms/step - loss: 0.4383 - accuracy: 0.7931 - auc: 0.8786 - val_loss: 0.5891 - val_accuracy: 0.6994 - val_auc: 0.7753 - lr: 1.0000e-04
Epoch 15/50
195/195 [==============================] - ETA: 0s - loss: 0.4430 - accuracy: 0.7909 - auc: 0.8743
Epoch 15: val_accuracy did not improve from 0.71484
195/195 [==============================] - 98s 504ms/step - loss: 0.4430 - accuracy: 0.7909 - auc: 0.8743 - val_loss: 0.5879 - val_accuracy: 0.7081 - val_auc: 0.7764 - lr: 1.0000e-04
Epoch 16/50
195/195 [==============================] - ETA: 0s - loss: 0.4465 - accuracy: 0.7825 - auc: 0.8705
Epoch 16: val_accuracy did not improve from 0.71484
195/195 [==============================] - 104s 532ms/step - loss: 0.4465 - accuracy: 0.7825 - auc: 0.8705 - val_loss: 0.6205 - val_accuracy: 0.6830 - val_auc: 0.7786 - lr: 1.0000e-04
Epoch 17/50
195/195 [==============================] - ETA: 0s - loss: 0.4437 - accuracy: 0.7925 - auc: 0.8741
Epoch 17: val_accuracy did not improve from 0.71484
195/195 [==============================] - 110s 566ms/step - loss: 0.4437 - accuracy: 0.7925 - auc: 0.8741 - val_loss: 0.6044 - val_accuracy: 0.6994 - val_auc: 0.7817 - lr: 1.0000e-04
Epoch 18/50
195/195 [==============================] - ETA: 0s - loss: 0.4224 - accuracy: 0.8050 - auc: 0.8896
Epoch 18: val_accuracy did not improve from 0.71484
195/195 [==============================] - 116s 596ms/step - loss: 0.4224 - accuracy: 0.8050 - auc: 0.8896 - val_loss: 0.5795 - val_accuracy: 0.7129 - val_auc: 0.7815 - lr: 1.0000e-05
Epoch 19/50
195/195 [==============================] - ETA: 0s - loss: 0.4248 - accuracy: 0.8028 - auc: 0.8875
Epoch 19: val_accuracy improved from 0.71484 to 0.71676, saving model to models_bkp\best_model5.h5
195/195 [==============================] - 115s 590ms/step - loss: 0.4248 - accuracy: 0.8028 - auc: 0.8875 - val_loss: 0.5790 - val_accuracy: 0.7168 - val_auc: 0.7818 - lr: 1.0000e-05
Epoch 20/50
195/195 [==============================] - ETA: 0s - loss: 0.4256 - accuracy: 0.8108 - auc: 0.8877
Epoch 20: val_accuracy improved from 0.71676 to 0.71965, saving model to models_bkp\best_model5.h5
195/195 [==============================] - 104s 534ms/step - loss: 0.4256 - accuracy: 0.8108 - auc: 0.8877 - val_loss: 0.5789 - val_accuracy: 0.7197 - val_auc: 0.7818 - lr: 1.0000e-05
Epoch 21/50
195/195 [==============================] - ETA: 0s - loss: 0.4296 - accuracy: 0.8130 - auc: 0.8863
Epoch 21: val_accuracy did not improve from 0.71965
195/195 [==============================] - 104s 531ms/step - loss: 0.4296 - accuracy: 0.8130 - auc: 0.8863 - val_loss: 0.5795 - val_accuracy: 0.7158 - val_auc: 0.7814 - lr: 1.0000e-05
Epoch 21: early stopping
In [550]:
model5 = load_model(model_path5)

Model5 Result¶

In [547]:
print_performance_results(model5)

Classes:  {'normal': 0, 'pneumonia': 1}

=================================================================

=================================================================
Accuracy:      61.52%
AUC:           71.00%
Precision:     44.58%
Recall:        38.70%
=================================================================
In [548]:
show_accuracy_charts(history5)
In [549]:
_CONFIDENCE_CRITERIA = .52
show_test_images_pred_img_panel(model5)

1/1 [==============================] - 0s 37ms/step
[[0.15321851]]
1/1 [==============================] - 0s 20ms/step
[[0.60776967]]
1/1 [==============================] - 0s 20ms/step
[[0.78351]]
1/1 [==============================] - 0s 30ms/step
[[0.05015829]]
1/1 [==============================] - 0s 23ms/step
[[0.36371034]]
1/1 [==============================] - 0s 20ms/step
[[0.06952075]]
1/1 [==============================] - 0s 18ms/step
[[0.448549]]
1/1 [==============================] - 0s 19ms/step
[[0.23729156]]
1/1 [==============================] - 0s 20ms/step
[[0.709824]]
1/1 [==============================] - 0s 15ms/step
[[0.38517132]]
1/1 [==============================] - 0s 12ms/step
[[0.38598502]]
1/1 [==============================] - 0s 16ms/step
[[0.7682301]]

Image Annotations using YOLOV5¶

In [465]:
df_data_final.head()
Out[465]:
level_0 patientId class x y width height Target Patient Sex Patient's Age View Position Image Size
0 9198 1aea68e3-a8b7-4402-92a5-16a1f23c6d4a Lung Opacity 676.0 262.0 202.0 453.0 1 M 41 AP 1024x1024
1 12539 ac056286-2e9a-43a7-8143-60082ef45a8b Normal NaN NaN NaN NaN 0 F 65 PA 1024x1024
2 6532 6a0ece7e-c22b-4e33-9544-3e26b624ef82 No Lung Opacity / Not Normal NaN NaN NaN NaN 0 M 35 AP 1024x1024
3 13470 ba6d52f7-d113-48aa-8adc-49f963ed8979 Normal NaN NaN NaN NaN 0 M 57 PA 1024x1024
4 7168 be00aef3-ebcf-4306-9960-a41b3c0cdafa Lung Opacity 621.0 234.0 188.0 271.0 1 M 64 PA 1024x1024
In [87]:
#copying the relavant information in our sample data for using it in faster rcnn
df_fcnn = df_data_final.loc[:,["patientId","x","y","width","height","Target"]]
In [88]:
mapping_target_info = {
    0:'normal',
    1:'pneumonia'
}
df_fcnn["ClassName"] = [mapping_target_info[trgt] for trgt in  df_fcnn["Target"]]
df_fcnn["patientId"] = df_fcnn["patientId"] + ".jpg"
In [89]:
df_fcnn.head()
Out[89]:
patientId x y width height Target ClassName
0 48a44f52-7fb3-4384-a76b-f6170b03a357.jpg NaN NaN NaN NaN 0 normal
1 49b95513-daab-49bb-bc6e-c5254ab1bc07.jpg 599.0 202.0 216.0 537.0 1 pneumonia
2 a985f470-12d9-48b0-bb1a-5bdffc657c47.jpg NaN NaN NaN NaN 0 normal
3 75ac9b0b-6f99-49f1-865c-3ccf54144b1a.jpg NaN NaN NaN NaN 0 normal
4 33c66fa9-6a08-4d0d-b309-88ce91a217c9.jpg 193.0 468.0 209.0 120.0 1 pneumonia
In [90]:
df_fcnn.rename(columns = {"patientId":"ImageID",'x':'XMin', 'y':'YMin','width':'XMax','height':'YMax'}, inplace = True)
df_fcnn['XMax'] = df_fcnn['XMin'] +  df_fcnn['XMax']
df_fcnn['YMax'] = df_fcnn['YMin'] +  df_fcnn['YMax']
df_fcnn['Xc'] = (df_fcnn['XMin'] +  df_fcnn['XMax'])/2
df_fcnn['Yc'] = (df_fcnn['YMin'] +  df_fcnn['YMax'])/2

# df_fcnn['XMin'] = df_fcnn['XMin']/1024
# df_fcnn['XMax'] = df_fcnn['XMax']/1024
# df_fcnn['YMin'] = df_fcnn['YMin']/1024
# df_fcnn['YMax'] = df_fcnn['YMax']/1024
print(df_fcnn.shape)
df_fcnn.head()

(6694, 9)
Out[90]:
ImageID XMin YMin XMax YMax Target ClassName Xc Yc
0 48a44f52-7fb3-4384-a76b-f6170b03a357.jpg NaN NaN NaN NaN 0 normal NaN NaN
1 49b95513-daab-49bb-bc6e-c5254ab1bc07.jpg 599.0 202.0 815.0 739.0 1 pneumonia 707.0 470.5
2 a985f470-12d9-48b0-bb1a-5bdffc657c47.jpg NaN NaN NaN NaN 0 normal NaN NaN
3 75ac9b0b-6f99-49f1-865c-3ccf54144b1a.jpg NaN NaN NaN NaN 0 normal NaN NaN
4 33c66fa9-6a08-4d0d-b309-88ce91a217c9.jpg 193.0 468.0 402.0 588.0 1 pneumonia 297.5 528.0
In [105]:
import torch
In [125]:
def create_YOLO_testtrainval_folder_structure():
    
    train_ = img_working_split_ + 'train/'
    test_ = img_working_split_ + 'test/'
    val_ = img_working_split_ + 'val/'
    
    _trNfiles = os.listdir(train_+'normal/')
    _trPfiles = os.listdir(train_+'pneumonia/')
    
    _tsNfiles = os.listdir(test_+'normal/')
    _tsPfiles = os.listdir(test_+'pneumonia/')
                           
    _vaNfiles = os.listdir(val_+'normal/')
    _vaPfiles = os.listdir(val_+'pneumonia/')
    
    
    """Train Annotations"""
    ii = 0
    
    x1,x2,y1,y2 = 0,0,0,0
    for _file in _trNfiles: #Train Normal
        _file_txt = _file.replace('.jpg','.txt')
        to_impath_ = os.path.join(img_working_yolo_images_train_dirname,_file) 
        dr_ = df_fcnn[df_fcnn["ImageID"] == _file]
        for idx, row in dr_.iterrows():
            cls_name = row["ClassName"]
            impath_ = train_+ cls_name+"/" + _file
        shutil.copy(impath_,to_impath_)
        ii += 1

    for _file in _trPfiles: #Train Pneumonia
        _file_txt = _file.replace('.jpg','.txt')
        to_impath_ = os.path.join(img_working_yolo_images_train_dirname,_file) 
        txtpath_ = os.path.join(img_working_yolo_labels_train_dirname,_file_txt)
        dr_ = df_fcnn[df_fcnn["ImageID"] == _file]
        for idx, row in dr_.iterrows():
            x1 = np.round(row["Xc"]/1024,5)
            y1 = np.round(row["Yc"]/1024,5)
            x2 = np.round(row["XMax"]/1024,5)
            y2 = np.round(row["YMax"]/1024,5)
            
            cls_name = row["ClassName"]
            impath_ = train_+ cls_name+"/" + _file
            with open(txtpath_,"w+") as f:
                f.write("1 " + str(x1) + " " + str(y1) + " " + str(x2) + " " + str(y2) + "\n") 
        shutil.copy(impath_,to_impath_)
        ii += 1
            
    print("Total Train text and images transfered to the respective YOLO directory : ", ii)
    
    
    """Test Annotations"""
    ii = 0
    x1,x2,y1,y2 = 0,0,0,0
    for _file in _tsNfiles: #Test Normal
        _file_txt = _file.replace('.jpg','.txt')
        dr_ = df_fcnn[df_fcnn["ImageID"] == _file]
        to_impath_ = os.path.join(img_working_yolo_images_test_dirname,_file) 
        for idx, row in dr_.iterrows():
            cls_name = row["ClassName"]
            impath_ = test_+ cls_name+"/" + _file
        shutil.copy(impath_,to_impath_)
        ii += 1

    for _file in _tsPfiles: #Test Pneumonia
        _file_txt = _file.replace('.jpg','.txt')
        dr_ = df_fcnn[df_fcnn["ImageID"] == _file]
        to_impath_ = os.path.join(img_working_yolo_images_test_dirname,_file) 
        txtpath_ = os.path.join(img_working_yolo_labels_test_dirname,_file_txt)
        for idx, row in dr_.iterrows():
            x1 = np.round(row["Xc"]/1024,5)
            y1 = np.round(row["Yc"]/1024,5)
            x2 = np.round(row["XMax"]/1024,5)
            y2 = np.round(row["YMax"]/1024,5)
            cls_name = row["ClassName"]
            impath_ = test_+ cls_name+"/" + _file
            with open(txtpath_,"w+") as f:
                f.write("1 " + str(x1) + " " + str(y1) + " " + str(x2) + " " + str(y2) + "\n") 
        shutil.copy(impath_,to_impath_)
        ii += 1
        
    print("Total Test text and images transfered to the respective YOLO directory : ", ii)
    
    """Validation Annotations"""
    ii = 0
    x1,x2,y1,y2 = 0,0,0,0
    for _file in _vaNfiles: #Normal
        _file_txt = _file.replace('.jpg','.txt')
        dr_ = df_fcnn[df_fcnn["ImageID"] == _file]
        to_impath_ = os.path.join(img_working_yolo_images_val_dirname,_file) 
        txtpath_ = os.path.join(img_working_yolo_labels_val_dirname,_file_txt)
        for idx, row in dr_.iterrows():
            cls_name = row["ClassName"]
            impath_ = val_+ cls_name+"/" + _file
        shutil.copy(impath_,to_impath_)
        ii += 1

    for _file in _vaPfiles: #Pneumonia
        _file_txt = _file.replace('.jpg','.txt')
        dr_ = df_fcnn[df_fcnn["ImageID"] == _file]
        to_impath_ = os.path.join(img_working_yolo_images_val_dirname,_file) 
        txtpath_ = os.path.join(img_working_yolo_labels_val_dirname,_file_txt)
        for idx, row in dr_.iterrows():
            x1 = np.round(row["Xc"]/1024,5)
            y1 = np.round(row["Yc"]/1024,5)
            x2 = np.round(row["XMax"]/1024,5)
            y2 = np.round(row["YMax"]/1024,5)
            cls_name = row["ClassName"]
            impath_ = val_+ cls_name+"/" + _file
            with open(txtpath_,"w+") as f:
                f.write("1 " + str(x1) + " " + str(y1) + " " + str(x2) + " " + str(y2) + "\n") 
        shutil.copy(impath_,to_impath_)
        ii += 1
        
    print("Total Val text and images transfered to the respective YOLO directory : ", ii)
In [126]:
#creating folder structure
create_image_directory_structure()
create_YOLO_testtrainval_folder_structure()

Total Train text and images transfered to the respective YOLO directory :  3131
Total Test text and images transfered to the respective YOLO directory :  1044
Total Val text and images transfered to the respective YOLO directory :  1044
In [1]:
#!git clone https://github.com/ultralytics/yolov5  # clone
%cd yolov5
#%pip install -qr requirements.txt  # install

import torch
import utils
display = utils.notebook_init()  # checks

YOLOv5  v7.0-63-gcdd804d Python-3.10.8 torch-1.13.1 CUDA:0 (NVIDIA GeForce RTX 3060 Laptop GPU, 6144MiB)

Setup complete  (16 CPUs, 15.8 GB RAM, 238.6/317.8 GB disk)
In [ ]:
# Train YOLOv5s on COCO128 for 3 epochs
!python train.py --img 1024 --batch 16 --epochs 30 --data yolo_images_data.yaml --weights yolov5s.pt --cache
In [ ]:
In [ ]:
In [ ]: